View
3
Download
0
Category
Preview:
Citation preview
Rewriting History * †
Alexander Ljungqvist Christopher Malloy
Stern School of Business London Business School
New York University
and CEPR
Felicia Marston
McIntire School of Commerce
University of Virginia
March 13, 2007
* Thanks for helpful comments go to Viral Acharya, Nick Barberis, Brad Barber, Lauren Cohen, Jennifer Juergens,
Jay Ritter, and Kent Womack, and to seminar participants at the U.S. Securities and Exchange Commission, the
2006 UNC-Duke Corporate Finance Conference, the 2007 AFA Conference in Chicago, Oppenheimer Funds,
Barclays Global Investors, the University of Sydney, the University of New South Wales, the University of
Auckland, Simon Fraser University, the University of Virginia, the University of Illinois, Dartmouth College, USC,
UCLA, Yale University, Tilburg University, and London Business School. We are grateful to Mark Chen for
sharing his data on Wall Street Journal “top stock pickers” and to Ruiming Lin, Pedro Saffi, and Yili Zhang for
excellent research assistance. We gratefully acknowledge the contribution of Thomson Financial for providing
broker recommendations data, available through the Institutional Brokers Estimate System. We are also grateful to
numerous analysts for patiently answering our questions. All errors are our own. † Address for correspondence: Salomon Center, Stern School of Business, New York University, Suite 9-160, 44
West Fourth Street, New York NY 10012-1126. Phone 212-998-0304. Fax 212-995-4220. e-mail
aljungqv@stern.nyu.edu.
2
Rewriting History
Abstract
Comparing two snapshots of the historical I/B/E/S database of research analyst stock
recommendations, taken in 2002 and 2004 but each covering the same time period 1993-2002,
we identify 54,729 ex post changes (out of 280,463 observations), including alterations of
recommendation levels, additions and deletions of records, and removal of analyst names. The
changes appear non-random across brokerage firms, analysts, and tickers, and have a significant
impact on the overall distribution of recommendations across stocks and within individual stocks
and brokerage firms. They also affect trading signal classifications, back-testing inferences, track
records of individual analysts, and models of analysts’ career outcomes in the three years
following the changes.
Key words: Security analysts; Stock recommendations; Global Settlement; Career concerns;
Forensic finance.
JEL classification: G21, G24, G28, J24, J44.
Comparing two snapshots of the entire historical I/B/E/S database of research analyst stock
recommendations, taken in 2002 and 2004 but each covering the same time period 1993-2002, we
identify tens of thousands of changes which collectively call into question the principle of
replicability of empirical research. The changes are of four types: 1) The non-random removal of
19,904 analyst names from historic recommendations (“anonymizations”); 2) the addition of 19,204
new records that were not previously part of the database; 3) the removal of 4,923 records that had
been in the data; and 4) alterations to 10,698 historical recommendation levels. In total, we
document 54,729 ex post changes to a database originally containing 280,463 observations.
Our main contribution is to document the characteristics and effects of these pervasive changes.
The academic literature on analyst stock recommendations, using I/B/E/S data, is truly vast: As of
December 12, 2006, Google Scholar identifies 565 articles and working papers using the keywords
“I/B/E/S”, “analysts”, and “recommendations”. Given this keen academic interest, as well as the
intense scrutiny that research analysts face in the marketplace and the growing popularity of trading
strategies based on analyst output, changes to the historical I/B/E/S database are of obvious interest
to academics and practitioners alike. We demonstrate that the changes have a significant effect on
the distribution of recommendations, both overall and for individual stocks and individual
brokerage firms. Equally important, they affect trading signal classifications, back-testing
inferences, the track records of individual analysts, and models of analysts’ career outcomes in the
years since the changes occurred. Regrettably, none of the changes can easily be “undone” by
researchers, which makes replicating extant studies difficult. Our findings thus have potentially
important ramifications for existing and future empirical studies of equity analysts.
Why do the historical data now look different than they once did? The contents of the database
changed at some point between September 2002 and May 2004, a period that not only coincided
with close scrutiny of Wall Street research by regulators, Congress, and the courts, but also saw a
2
substantial downsizing of research departments at most major brokerage firms in the U.S.
According to communication received from Thomson Financial (the owner of I/B/E/S) in
November and December 2006, the anonymizations were caused by a series of software glitches,
introduced in 2002-2003.1 Surprisingly, despite the seemingly random nature of this type of shock
to the data, the resulting patterns have apparently systematic components, rather than appearing
random. For instance, bolder recommendations are more likely to be anonymized, as are
recommendations from more senior analysts and Institutional Investor “all-stars.” The
characteristics of the additions and deletions are similarly unusual. Additions disproportionately
consist of holds and sells; indeed, in the case of one prominent brokerage firm, 91.5% of its 234
additions are sells, and these increase the number of sells the firm has on the 2002 tape by a factor
of 20. Deletions, on the other hand, disproportionately consist of strong buys, while alterations
disproportionately consist of buys and strong buys (which are typically revised down). Perhaps
most strikingly, all four types of changes correlate strongly with survival by both the brokerage firm
and by the analyst.
We divide our analysis into two broad areas of interest: Individual recommendations and
analyst-level research. At the recommendation level, the collective effect of the changes is to make
the overall distribution of historic recommendations issued for U.S. companies between October
1993 and July 2002 appear more conservative in 2004 than it did in 2002, with considerably fewer
strong buys and buys and more holds, sells, and strong sells.2 The magnitude of this shift, while
1 As of February 2007, Thomson Financial’s position is that “changes to Recommendation History have occurred as
part of necessary processes and maintenance.” Previously, Thomson had indicated that the anonymizations were the
result of three software glitches. The first produced incorrect start and end dates for some analysts’ employment
histories. The second caused selective ticker histories for certain analysts to be anonymized, in connection with attempts
to consolidate duplicate analyst identifiers. The third, which can still occur, results in recommendations being attributed
to an analyst who has departed; subsequent corrections involve anonymizing such recommendations. In unreported tests
we find that these explanations appear to account for at most 70% of the anonymizations, but it is possible that other, as
yet unidentified, software glitches may account for the remainder. 2 This finding is distinct from Barber et al. (2006), who document a shift in the overall distribution of recommendations
over time; we focus on a constant time period, 1993 through 2002.
3
statistically significant, is economically small overall, because in any given year no more than a
sixth to a third of stocks are affected. For the subset of affected stocks, however, the changes result
in a decidedly more conservative distribution of recommendations, especially in the late 1990s – a
period that continues to be of central interest to researchers. Broker-level distributions, which can
be used to predict the profitability of stock recommendations (Barber et al. (2006)), change even
more dramatically. One reason why these findings are important is that the apparently optimistic
bias in the historical distribution of recommendations was a frequently cited impetus to the
regulatory proceedings that culminated in the Global Research Analyst Settlement of 2003 between
New York attorney general Eliot Spitzer, the SEC, NASD, NYSE, North American Securities
Administrators Association, state regulators, and twelve brokerage firms.
The changes to the I/B/E/S database also affect the classification of trading signals such as
“upgrades” and “downgrades,” the key inputs for a large literature on the profitability of analyst
recommendations (e.g., Womack (1996), Barber et al. (2001), or Jegadeesh et al. (2004); see
Ramnath, Rock, and Shane (2005) for a survey). Overall, using the 2004 tape instead of the 2002
tape results in 6.2% of upgrades and 5.3% of downgrades being “re-classified.” In addition, there is
a curious time trend, in that the alterations, deletions, and additions appear to affect more recent
recommendations, and especially those issued in 2002, 11% of which would be re-classified.
Similarly, the changes impact the implementation of popular trading strategies based on analyst
consensus recommendations. For example, employing a standard portfolio classification technique
that forms quintiles based on the quarterly change in consensus analyst recommendations (a robust
predictor of returns, as shown in Jegadeesh et al. (2004)) requires one to re-classify more than 17%
of the ticker-quarter observations when using the 2004 tape compared to the 2002 tape. A simple
trading strategy that buys stocks in the top consensus change quintile and shorts stocks in the lowest
quintile each month performs 15.9% to 42.4% better on the 2004 tape than on the 2002 tape,
4
suggesting that the changes to the I/B/E/S database have resulted in the appearance of more
profitable analyst research.
Track records of individual analysts are also affected. Analysts’ track records are the key
variable of interest in a large and growing literature on career concerns and agency problems in the
analyst industry. (In December 2006, Google Scholar lists 129 articles and working papers
containing the key words “analysts”, “conflicts of interest”, and “I/B/E/S”.) Among the group of
analysts with anonymized recommendations, we find that abnormal returns following subsequently
anonymized upgrades-to-buy are significantly lower than are abnormal returns following upgrades
by the same analysts that remained untouched. The return differential is large, ranging from 3.6% to
4.0% p.a. over the 1993-2002 period. It is even larger in the most recent, post-bubble period (as
high as 7.4% p.a.), and for the sub-sample of bolder recommendations (5.2% and 9.1% p.a. in the
full-sample and post-bubble periods, respectively).
Analysts whose track records are affected are associated with more favorable career outcomes
over the 2003-2005 period than their track records and abilities would otherwise warrant. Following
the career path modeling in Hong, Kubik, and Solomon (2000) closely, we find that analysts
associated with anonymizations experience a more than 60% increase in the likelihood of
subsequently moving from a low-status to a high-status brokerage firm. This effect is much larger
than any other in our career outcome models. The magnitude of the effect of additions on career
outcomes is also large, boosting the likelihood by more than 40%. Similarly, analysts are more
likely to be rated the top stock picker in their sectors by the Wall Street Journal, which relies on
Thomson Financial data, if some of their recommendations have been dropped or added.
5
Collectively, our findings raise serious doubts about the replicability of past, current, and future
studies using the I/B/E/S historical recommendations database.3
I. Data
I.A Sample Construction and Overview of Changes
The I/B/E/S historical detail recommendations database contains investment ratings for U.S.
listed companies issued by most of the brokerage firms active in the U.S. Our analysis compares
two snapshots of the same historical time period – October 29, 1993 to July 18, 2002 – taken at two
different points in time: September 2002 and May 2004. We will refer to these snapshots as the
2002 and 2004 tapes, respectively. (We ignore recommendations on the 2004 tape dated after July
18, 2002.) To meet the minimum scholarly standard that research findings be replicable, each new
tape must contain identical historical data (except for cleanups). This is not the case.
A typical I/B/E/S record includes the analyst’s name and her six-digit amaskcd identifier as
assigned by I/B/E/S; the name of the analyst’s employer at the time of the recommendation; the
I/B/E/S ticker and historical CUSIP of the company concerned; the date the recommendation was
issued; and the recommendation itself. Different brokerage firms use different wordings for their
recommendations, which I/B/E/S translates into a numerical score on the following scale: Strong
buy=1, buy=2, hold=3, sell=4, strong sell=5.
We merge the 2002 and 2004 tapes by I/B/E/S ticker,4 standardized brokerage firm name,5 and
date. This reveals four types of changes to the I/B/E/S historical recommendations database:
1) Alterations: Records whose recommendation levels are different on the 2002 and 2004 tapes;
3 For related work on the impact of problems in other archival databases (e.g., CRSP, Compustat), see Rosenberg and
Houglet (1974), Bennin (1980), Shumway (1997), Canina et al. (1998), Shumway and Warther (1999), and Elton,
Gruber, and Blake (2001). See http://www.kellogg.northwestern.edu/rc/crsp-cstat-references.htm for a summary. 4 We have verified that the changes we document are not due to changing I/B/E/S tickers.
5 In some cases, I/B/E/S uses multiple codes to identify the same brokerage firm (e.g., NOMURA and NOMURAUS
both decode to Nomura Securities). We standardize such name variations.
6
2) Deletions: Records that appear on the 2002 tape but not on the 2004 tape;
3) Additions: Records that appear on the 2004 tape but not on the 2002 tape;6
4) Anonymizations: Records attributed to specific analysts on the 2002 tape whose names are
missing on the 2004 tape (and whose amaskcd identifiers have been set to zero).
There are 280,463 investment recommendations on the 2002 tape. By 2004, there are 10,698
alterations (3.8%), 4,923 deletions (1.8%), 19,204 additions (6.8%), and 19,904 anonymizations
(7.1%). We find very little overlap between the changes, suggesting they are almost always
independent of each other. For instance, it is rarely the case that an anonymized recommendation is
also altered, or that deletions are linked to additions. This makes it unlikely that I/B/E/S simply
dropped and replaced records containing errors.7 Overall, there are around 55,000 changes in total.
They are thus pervasive. Since 2004, there have been few additional alterations, additions, or
anonymizations, though deletions continue to occur; see Appendix A.8
Panel A of Table I presents a breakdown of each type of change, by year. Each is observed
throughout the 1993-2002 sample period, with no apparent time trends. The annual number of
affected records varies from 3,349 in 1993 to 6,929 in 1999.
Panel B of Table I presents a breakdown of each type of change, by I/B/E/S rating level. Of the
280,463 original recommendations on the 2002 tape, 28.6% are strong buys, 36.7% are buys, 31.4%
are holds, 1.7% are sells, and 1.6% are strong sells. The average recommendation level is 2.11. By
6 We exclude from our analysis 22,240 additions by brokerage firms that did not contribute data to I/B/E/S in 2002 but
appeared on the 2004 tape, as they have since disappeared. On closer inspection, these brokers are quantitative research
shops that produce algorithmic recommendations constrained to be symmetrically distributed. Any academic research
that inadvertently included their data will severely understate the well-known optimistic bias in recommendations. 7 Specifically, when we form all pairwise matches of additions and deletions on {broker, recommendation year, I/B/E/S
ticker}, we find only 301 pairwise matches. Likewise, there are few (922) pairwise matches of additions and alterations,
and even fewer (25) of alterations and deletions. 8 Using earlier and later tapes, we can document the time series of changes. For instance, between January and
September 2002, there were nine anonymizations, i.e. one per month on average. Between September 2002 and May
2004, the sample period which we focus on, there were 995 anonymizations per month on average. Between May 2004
and February 2006, there were 591 anonymizations, or 28.1 per month on average.
7
comparison, the 10,698 records subject to alteration have an average recommendation of 1.98 on the
2002 tape and include disproportionately few holds. On the 2004 tape, their average is more
pessimistic, largely because of the near trebling in the number of negative recommendations.
Deletions are disproportionately strong buys, with an average recommendation level of 1.7.
Additions are the opposite, containing disproportionately holds and sells, with an average
recommendation level of 2.44. The distribution of the 19,904 anonymized recommendations largely
mirrors that of the 2002 tape. As we show in Section II.A, the net effect of alterations, deletions,
and additions is to make the distribution of recommendations over the 1993-2002 time period
appear more conservative.
Panel C of Table I reveals an interesting time trend among the alterations. In the early sample
years, alterations result in about as many downward as upward revisions of the original ratings.
Since the end of the 1990s boom market, on the other hand, and especially in 2002, negative
revisions significantly outnumber positive ones. Thus, alterations contribute to the appearance of
more conservative sell-side research in the early 2000s, a period of generally falling share prices.
I.B Characteristics of Ex Post Changes
Affected records have certain characteristics in common that appear non-random. In Table II,
we split brokerage firms into two groups depending on whether any of their recommendations have
been altered, deleted, added to, or anonymized.9 Around one in six brokerage firms is associated
with alterations, one in three with deletions and with additions, and nearly two in three with
anonymizations.10
9 Many brokerage firms have merged over the sample period. After allocating the historic recommendations of the
predecessors to the surviving brokerage firm as of June 30, 2002 using the bank merger data described in Asker and
Ljungqvist (2005), there are 385 distinct brokerage firms. Sixteen firms never revealed analyst names on the 2002 tape,
leaving 369 brokerage firms whose recommendations could be subject to anonymization. 10 We obtain qualitatively similar results if we require a substantial fraction of a brokerage firm’s recommendations
(such as 5%) to have been affected, rather than “at least one.”
8
Brokerage firms associated with these changes are substantially larger, both in terms of the size
of their investment banking operations and the size of their research departments. Most remarkably,
they employ between eight and 16 times more analysts on average in 2002 than do unaffected
brokerage firms. The main cause of this difference is the fact that the group of unaffected firms
includes a disproportionate number with zero headcounts in 2002, suggesting they have ceased
publishing sell-side research or at least stopped contributing data to I/B/E/S.11 In fact, continuing to
publish research appears to be a pre-condition for a brokerage firm’s recommendations to have
changed: Not a single recommendation associated with one of the 89 brokerage firms that have
ceased publishing investment research by 2002 has been anonymized, whereas an astonishing
85.4% of the 280 firms that continue publishing research had some recommendations anonymized.
Similar, though slightly less extreme patterns hold for alterations, deletions, and additions.
Beyond size and continued production of sell-side research, affected brokerage firms have two
further characteristics in common: They include the twelve firms sanctioned in the Global
Settlement (except for alterations) and they tend to be integrated securities firms, offering both
investment banking and research services. For example, of the 65 brokerage firms that lead-
managed at least one underwritten equity transaction in 2002 (our proxy for the presence of an
investment banking operation), 61 (93.8%) saw some anonymizations, whereas of the 304
brokerage firms without investment banking operations in 2002, only 178 (58.6%) did.
We find similarly strong and systematic patterns at the analyst level. Table III splits analysts
into two groups depending on whether any of their recommendations have been altered, deleted,
added to, or anonymized. Of the 7,817 analysts with named records on the 2002 tape, 12% are
associated with alterations, 11.1% with deletions, 23.6% with additions, and 27.3% with
anonymizations. The average number of affected records per analyst varies from 5.5 deletions to
11 Brokerage firms that have ceased publishing research are distinct from those taken over by another broker.
9
10.9 alterations (though the medians are lower for each type of change). This corresponds to
between 17.6% and 40.2% of the average affected analyst’s historical records on the 2002 tape.
Panel B shows that analysts associated with alterations, deletions, or additions work for
significantly smaller brokerage firms on average, and analysts at sanctioned brokers are less likely
to have records altered or deleted on the 2004 tape.
According to Panel C, affected analysts appear to have the very best job prospects (Hong and
Kubik (2003)). They are not only more likely to be Institutional Investor “all-stars” or “top stock
pickers” according to the Wall Street Journal “Best on the Street” survey, they are also associated
with more accurate earnings forecasts compared to their sector peers, as defined in Hong, Kubik,
and Solomon (2000),12 and tend to be more senior as judged by the average time they have been in
the profession.13 Finally, they issue significantly more recommendations than unaffected analysts.14
Arguably the most remarkable result of Table III is shown in Panel D. Much like all four types
of change correlate with survival by the brokerage firm, we find that they also correlate with
survival by the analyst: Affected analysts are vastly more likely to remain in the industry beyond
2002. For example, 70.8% of the anonymizers continue to contribute recommendations and
earnings forecasts to I/B/E/S after 2002, compared to only 46.9% of the non-anonymizers.15
In Table IV, we ask whether conditional on brokerage firm and analyst characteristics,
12 The absolute forecast error of analyst i covering company k in year t is the difference between the analyst’s most
recent forecast of year-end EPS (issued between January 1 and July 1 of that year) and subsequent realized earnings.
Absolute forecast errors are scaled so that the most accurate analyst scores one and the least accurate zero. The analyst’s
relative forecast accuracy in year t is her average score across the k stocks she covers over years t-2 to t. 13 We measure analyst characteristics as of the earlier of 2002 or the date of the analyst’s retirement.
14 Analysts with these characteristics tend to publish more, and Table III shows that more prolific analysts are more
likely to be associated with affected records. However, this correlation does not drive the summary statistics. For
example, we find the same patterns if we focus on analysts who make an above-median number of recommendations. 15 Because we define continued employment in the industry based on an analyst contributing data to I/B/E/S, these
numbers are lower bounds. Analysts whose employers cease to contribute data to I/B/E/S (e.g., S&P), or those moving
to employers that do not contribute data to I/B/E/S, will be classified as having left the industry, as will analysts who no
longer have their name listed first on research reports (perhaps because a more senior analyst has been recruited to the
team) or who have been promoted to the position of research director (and so stop publishing).
10
recommendations have changed randomly. The unit of analysis here is thus a recommendation,
rather than a brokerage firm or an analyst. We use multivariate probit models to relate the
probability that a given recommendation is altered, dropped, added, or anonymized to the
characteristics of the analyst and her brokerage firm, three recommendation-level attributes, as well
as random brokerage effects to control for otherwise omitted heterogeneity arising from differences
across brokerage firms.
Judged by the pseudo-R2, which range from 47.3% to 75.8%, each type of change appears
highly predictable rather than random. All else equal, deletions, additions, and anonymizations are
25%, 20%, and 117% more likely if the analyst continues to be employed in the industry. By
implication, the conditional likelihood of a recommendation made by an analyst who has retired by
2002 being anonymized is zero. The t-statistic for this variable is 32.1, making it easily the most
significant determinant of anonymization.16
Recommendations from Institutional Investor all-stars are 21% less likely to be altered, 10%
less likely to be dropped, 10% less likely to be added, and 61% more likely to be anonymized.
Those from Wall Street Journal top stock pickers are a third more likely to have been added or
anonymized. Indeed, with the exception of all-star status, additions and anonymizations share many
of the same determinants; for instance, they are more likely among analysts with fewer
recommendations to their name. Seniority has a positive effect on the likelihood of each type of
change, but the economic magnitudes are more modest.
Unlike the size of the brokerage firm’s investment banking operation, the size of the research
department has a large effect: Recommendations are 13% more likely to be altered, 48% more
likely to be deleted, 19% more likely to be added, and 41% more likely to be anonymized for a one
16 Results are robust to excluding from the set of anonymizers analysts with very few or very many anonymizations.
11
standard deviation increase in the log number of analysts employed in 2002. Alterations and
additions are significantly less likely to occur at sanctioned brokerage firms.
Whether the brokerage firm has stopped contributing data to I/B/E/S predicts anonymization
perfectly and so cannot be controlled for in that model. We also see significantly fewer alterations
and additions at inactive brokerage firms. Deletions, on the other hand, are 143% more likely if the
brokerage firm no longer contributes data to I/B/E/S. (We find no support for the hypothesis that
such brokerage firms requested all their historical data to be removed from I/B/E/S.)
Among the recommendation-level characteristics, bolder recommendations are more likely to be
altered or anonymized, and less likely to be dropped. As we will see in Section III, bold
anonymized recommendations tend to be poor performers. Finally, additions are less likely for
investment banking clients or large issuers of securities.
In sum, the changes to the I/B/E/S historical recommendations database are widespread and
appear non-random across brokerage firms, analysts, and tickers.
I.C Do the Changes Reflect Data Corrections?
It is possible that records changed simply because I/B/E/S discovered errors on the 2002 tape.
To rule out this possibility, we hand-check the data integrity of the 2002 tape against the analyst
report collection available through Thomson’s Investext service and news sources accessed through
the Reuters/Dow Jones Factiva service.
In order to investigate the anonymizations, we focus on six randomly selected brokerage firms,
namely three bulge-bracket investment banks and three medium-sized brokerage firms. (The terms
governing our use of the I/B/E/S data prevent us from naming the firms we selected.) We are able to
verify the analyst’s identity in 1,144 anonymization cases. In only 43 of these (3.8%) does the name
of the analyst writing the report not match the name on the 2002 I/B/E/S tape, indicating that the
names on the original 2002 tape are largely accurate.
12
We conduct a series of similar, random Investext searches for the three other types of changes.17
For the deletions, we are able to identify 96 exact matches by broker/ticker/date, only two of which
suggest that the analyst’s name on the 2002 tape was incorrect. Widening the match window to two
weeks on either side of the I/B/E/S recommendation date yields 147 additional matches, of which
only three indicate an obvious error in the original 2002 record. Thus, at least for this small, random
sample, the vast majority of the deleted records (97.9%) appear to be valid recommendations that
should not have been deleted.
We are able find less than 20% of the 1,055 additions we hand-checked in Investext, even if we
widen the window to two weeks either side. This could mean that some additions are not bona fide
historical recommendations, or if they were, that they were not available through public channels
such as I/B/E/S and Investext. Moreover, as the origin of the additions is unclear, it is possible that
the 19,204 additions are not a random subset of the bona fide historical recommendations that were
for some reason not on the 2002 tape. We can, however, rule out two possibilities: 1) That the
additions are the result of merging I/B/E/S and First Call (another Thomson Financial database of
stock recommendations),18 and 2) that the additions represent a simple expansion of the I/B/E/S
coverage (i.e., the ex post inclusion of new analysts, new brokers, or new tickers).19
Alterations include many cases where every historical buy (say) at a given brokerage firm was
recoded as a strong buy. It is possible that this type of case reflects the erroneous retroactive
application of a broker’s new rating scale to its historical data. Kadan et al. (2005) show that many
17 Thomson Financial claims that the causal factors for the additions, alterations, and deletions are “contributor rating
scale changes, contributor code mergers, and the removal of Rankings Only contributors.” As detailed in Table VIII
(and in similar unreported tests at the broker level) and in the Appendix, these factors do not appear to fully explain the
patterns we see. Moreover, they do not appear to address additions. 18 Only 6,295 of the 19,204 additions (32.8%) can be found in First Call when matching on standardized broker name,
historical CUSIP, and date (allowing for a two-week window either side of the I/B/E/S date). The match rate is similarly
low for the entire 2002 I/B/E/S tape: Only 46.8% of the 280,463 I/B/E/S recommendations can be found in First Call. 19 Only 1,033 out of the 19,204 additions (5.4%) correspond to an expansion of historical coverage in the database.
13
brokerage firms adopted a three-point rating scale at various points in 2002 and 2003, though there
is no logical reason why they should have applied their new scales to their historical I/B/E/S
records. In many other cases, some but not all records with the same rating are recoded, indicating
other reasons for the alterations than the adoption of a new rating scale. Often, alterations apply not
just to the I/B/E/S rating code (“strong buy” and so on), but also to the original broker rating (“long
term buy” and so on). Our Investext analysis identifies 130 exact matches by broker/ticker/date,
none of which appears to be a valid data correction.
Overall, our random searches suggest that anonymizations, alterations, and deletions are likely
not simply the result of attempts to correct data errors on the 2002 tape, while the status of additions
remains unclear.
I.D Can the Changes Easily Be Undone?
It is impossible for users to undo the deletions or additions to the I/B/E/S historical
recommendations database without access to a historic snapshot of the data, such as our 2002 tape.
In theory, there are ways to potentially undo the alterations, but as we argue more fully in Appendix
A, such fixes are incomplete and can introduce errors into the data. In response to this paper,
Thomson Financial have recently reinstated the missing analyst names in the recommendation
history file; as of February 12, 2007, the data on WRDS now reflect this reinstatement. Again in
theory, users of downloads/snapshots of the recommendations database from the period May 2004
to December 2006 can repopulate the missing analyst names from the I/B/E/S earnings forecast
database, but as described in Appendix A, this procedure is problematic. Furthermore, regardless of
whether the changes can or will be undone, they have affected research conducted over the last
three or four years, as shown in the next two sections.
II. The Impact of the Data Changes on Recommendation-level Analysis
It is difficult to overstate the significance of the changes to the I/B/E/S historical
14
recommendations database outlined in the previous section. The alterations, deletions, and additions
combine to make the industry-wide distribution of recommendations, the recommendation
distributions for individual stocks, and those for individual brokerage firms look considerably more
conservative than they did according to the 2002 tape. They also affect historical “trading signals”
such as upgrades and downgrades as well as the level of the historical consensus. Each of these is at
the heart of a separate line of research. In this section, we document the potential effects of the
I/B/E/S changes for research, while bearing in mind that they may also affect the work of
regulators, legislators, litigators, and investment professionals, all of whom rely on archival
databases such as I/B/E/S.
II.A Effects on the Distribution of Stock Recommendations
As we saw in Section I, the distributions of altered, dropped, and added recommendations are
highly unusual: Altered records are disproportionately optimistic on the 2002 tape and become more
pessimistic on the 2004 tape; dropped records too are unusually optimistic; whereas added records
are uncommonly pessimistic. Panel A of Table V shows the net effect of these changes on the
overall distribution of historic stock recommendations over the 1993-2002 period. Overall, the
changes make the distribution appear more conservative in 2004 than it did in 2002. The net
difference between the two tapes is heavily skewed towards holds, sells, and strong sells, which
collectively account for 82.4% of the net changes; for comparison, pessimistic ratings account for
only 34.7% of the original 2002 tape.
Panel A understates the effect of the changes on the overall distribution, because the effect turns
out to be concentrated in a relatively small number of stocks. Panel B focuses on changes in annual
stock-level distributions. The 2002 tape contains 49,097 ticker-years, with an average of 5.7
recommendations per ticker and year. Overall, as in Panel A, we see that the distribution of
recommendations for the average ticker-year has clearly become more conservative in the wake of
15
the alterations, deletions, and additions: For the average ticker, the proportion of buys (including
strong buys) decreases by 1.0 percentage points, while the proportion of holds and sells (including
strong sells) increases by 0.4 and 0.6 percentage points, respectively. But the majority of tickers
experience no change in their recommendation distribution. For the subset of stocks that do, the
shift is large: The 24% of the total stock universe covered by analysts in I/B/E/S that are affected by
the changes now show a decidedly more conservative distribution of recommendations over the
1993-2002 period. For example, at the height of the bull market in 2000, the proportion of buys and
strong buys for the average affected stock decreases by more than six percentage points, from
68.0% to 61.7%, with a corresponding increase in holds and, especially, sells.
Barber et al. (2006) show that broker-level recommendation distributions can be used to predict
the profitability of stock recommendations. The underlying idea is that a strong buy from a
brokerage firm that relatively rarely issues strong buys is more informative, and hence more
profitable to trade on, than a strong buy from a generally more optimistic broker. As Panel C of
Table VI shows, broker-level recommendation distributions also change significantly as a result of
the alterations, deletions, and additions to the I/B/E/S historical recommendations database.
Following Barber et al. (2006), we rank brokerage firms every quarter based on the fraction of
their end-of-quarter outstanding recommendations that are buys or strong buys and group them into
quintiles, using data from either the 2002 tape or the 2004 tape. Overall, 1,194 of the 6,540 broker-
quarters (18.3%) are assigned to different quintiles on the two tapes. For instance, 18.7% of the
most bullish brokers on the 2002 tape migrate into another quintile on the 2004 tape. Clearly, these
migrations, caused by the changes to the I/B/E/S historical recommendations database, have the
potential to affect the replicability of broker-level research.
II.B Effects on Trading Signals
As important as the effects on the overall, stock-level, and broker-level distributions of
16
recommendations levels are the effects of the alterations, deletions, and additions on the distribution
of recommendation changes or “trading signals”, the key inputs for a vast literature on the
profitability of analyst recommendations (see Ramnath, Rock, and Shane (2005) for a review). We
construct trading signals for broker/ticker pairs using a twelve-month look-back window. For
instance, a downgrade is defined as a negative change from a recommendation issued by the same
broker for the same I/B/E/S ticker within the previous twelve months. If the previous
recommendation was issued more than twelve months ago, the current recommendation is defined
to be a reinitiation. In the absence of a prior recommendation, the current recommendation is
defined to be an initiation.
Panel A of Table VI provides a breakdown of the distribution of trading signals for the 2002
tape, each of the three types of changes, and the resulting 2004 tape. The effect of the alterations is
to increase the number of reiterations at the expense of upgrades and downgrades, blunting the
original trading signals. Deleted records were disproportionately upgrades, while added records are
disproportionately downgrades. The combined effect of the changes is to increase the fraction of
downgrades and reiterations on the 2004 tape.20 While not shown, the effect becomes even more
pronounced once we exclude unaffected stocks, as in the previous table.
Panel B provides a transition matrix for the changed trading signals from the 2002 to the 2004
tape. Of the 56,464 records originally classified as downgrades on the 2002 tape, 2,981 (5.3%) are
classified differently on the 2004 tape: 137 now appear as upgrades, 1,797 as reiterations, 263 as
initiations, and 181 as reinitiations, while 603 have been deleted. Similarly, 6.2% of the original
upgrades, 11% of the original reiterations, 4.7% of the original initiations, and 4.8% of the original
reinitiations have migrated to a different trading signal category on the 2004 tape. When we take the
20 The net number of changed trading signals is less than the sum of the trading signals that are changed due to
alterations, deletions, and additions, as some of these changes cancel each other out.
17
additions into account, as many as 34,804 (12.4%) of the original trading signals for the 1993-2002
period have changed or been dropped on the 2004 tape.
Figure 1 reveals an interesting time trend to the trading signal changes, by plotting the fraction
of affected records per year. For each type of trading signal, the alterations, deletions, and additions
have the largest effect among the more recent recommendations, and especially among those issued
in 2002: 7.8% of original year-2002 downgrades have changed, as have 12.3% of upgrades, 21.7%
of reiterations, 13.1% of initiations, and 5.5% of reinitiations. On net, 11% of year-2002 trading
signals are classified differently on the 2002 and 2004 tapes.21
II.C Effects on Consensus Recommendations
Consensus recommendations are commonly employed in quantitative trading strategies,
following evidence that sorting based on consensus recommendations (Barber et al. (2001)), and
particularly on changes in consensus recommendations (Jegadeesh et al. (2004)), is a profitable
strategy. We examine the impact of the alterations, deletions, and additions on a popular strategy
that trades on changes in consensus recommendations in the pre-2000 period, as Barber et al. (2003)
show that strategies based on consensus recommendations perform poorly after 2000.22 Our goal is
to assess whether the changes to the I/B/E/S database affect back-tests of a key stylized fact in the
literature; it is not to question the validity of any particular finding.
We employ a standard portfolio classification technique that forms quintiles each month based
on the lagged quarterly change in consensus recommendations. Recommendations are reverse-
scored from 5 (strong buy) to 1 (strong sell). The consensus recommendation for a ticker equals the
mean outstanding recommendation at the end of a calendar quarter, based on a minimum of three
recommendations. Firms are grouped into quintiles at the beginning of the next quarter based on the
21 These numbers are based on our standardized brokerage firm names. If we use I/B/E/S’s own brokerage firm names
(i.e., “brkcd”), the fraction of misclassified trading signals increases; see the final graph of Figure 1. 22 We have verified this fact in unreported results.
18
change in consensus. Panel A of Table VII reports summary statistics on the consensus change in
each quintile using the 2002 I/B/E/S tape. Over the 1993-1999 period, analysts were slightly more
likely to downgrade a firm than upgrade it, as the mean quarterly consensus change equals -0.02.
Panels B and C illustrate the effect of using the 2004 tape instead of the 2002 tape. Panel B
reports the fraction of ticker/quarter observations within each quintile (and for the full sample) that
are classified in a different quintile when using the 2004 tape. On average, a striking 17.4% of the
total ticker/quarter observations, including 13% of the observations in quintiles 1 and 5, migrate to a
different quintile on the 2004 tape.
Panel C reports monthly portfolio returns for a simple trading strategy (“Spread”) that each
month buys stocks in the highest quarterly change quintile (Q5) and shorts stocks in the lowest
quarterly change quintile (Q1), with a one-month holding period. We calculate abnormal portfolio
returns in two ways: 1) By estimating a “four-factor” alpha (as in Carhart (1997)), and 2) by
estimating a “four-factor plus industry” alpha. Four-factor alpha returns ( jα ) are computed from
the following monthly time-series regression for each portfolio j:
,)( jttjtjtjftmtjjft
j
t UMDuHMLhSMBsRRRR εβα ++++−+=− (1)
where j
tR is the month t return on portfolio j, ftR is the month t risk-free rate, mtR is the month t
return on the CRSP value-weighted index, SMBt is the month t return on a value-weighted portfolio
of small stocks minus the month t return on a value-weighted portfolio of big stocks, HMLt is the
month t return on a value-weighted portfolio of high book-to-market stocks minus the month t
return on a value-weighted portfolio of stocks with low book-to-market, and UMDt is the month t
return on a value-weighted portfolio of stocks with high returns from month t-12 to month t-2 minus
the month t return on a value-weighted portfolio of stocks with low returns from month t-12 to
month t-2. The “4-factor plus industry alpha” is the intercept from a regression of the monthly
19
portfolio spread return on the four factors above plus industry excess returns (value-weighted excess
returns for each of ten industry segments as constructed by Kenneth French).
We execute the strategy identically on the 2002 tape and the 2004 tape, and report the
differences in results for the two tapes in Panel C. Consistent with prior evidence, the “spread”
strategy is a profitable one up until 2000, on both the 2002 and 2004 tapes. Interestingly, however,
the strategy performs significantly better on the 2004 tape than on the 2002 tape, despite the fact
that the strategy is back-tested here over the exact same time-period. The magnitude of this
difference is large: Between 14 and 20 basis points a month, which translates into an improvement
of 15.9% to 42.4% on the 2004 tape relative to the performance found on the 2002 tape.
In unreported tests, we find that most of the apparent improvement in the “spread” strategy is
due to the additions. Thus, a hypothetical investor who had access to the additions in real time
would have outperformed a hypothetical investor who only relied on publicly available
recommendations data, such as the I/B/E/S historical recommendations database.
III. The Impact of the Data Changes on Analyst-level Analysis
Recent research highlights potential conflicts of interest in the production of investment
research due to the competing roles sell-side analysts play in financial markets. Analysts may
respond to institutional incentives such as pressure from investment bankers (Lin and McNichols
(1998), Michaely and Womack (1999), Hong and Kubik (2003)), a desire to maintain ties to senior
management at the companies they cover (Francis and Philbrick (1993), Lim (2001)), or a desire to
generate commissions (Irvine (2004), Jackson (2005)). What is usually cited as the mechanism to
keep analysts in check is their incentive to maintain their reputations for accuracy, objectivity, and
independence (Stickel (1992), Hong, Kubik, and Salomon (2000), Ljungqvist et al. (2006)).
Reputable analysts generate more commission income for their employers (Irvine (2004), Cowen,
Groysberg, and Healy (2003), Jackson (2005)), receive higher compensation (Kothari (2001)), and
20
are more likely to be hired by the most prestigious brokerage firms (Hong and Kubik (2003), Hong,
Kubik, and Solomon (2000)). Poorly performing analysts, on the other hand, generate less
commission income and are more likely to lose their jobs (Mikhail, Walther, and Willis (1999)).
Ultimately, an analyst’s reputation is informed by her track record. In this section we investigate
the impact of the changes to the I/B/E/S historical recommendations database on the track records
and career outcomes of individual analysts.
III.A Pattern Analysis
The alterations, deletions, additions, and anonymizations form a diverse set of patterns at the
analyst level, reported in Table VIII. We define a “history” as a sequence of recommendations by
analyst i for ticker k, ordered in calendar time. The most common pattern – accounting for 32% of
the 10,698 alterations, 32% of the 4,923 deletions, and 40.4% of the 19,204 additions – is the
selective change of individual records within a given history. For instance, we might find that the
third and seventh recommendation in a history have been altered, or that the penultimate record has
been dropped. By contrast, it is relatively rare that an entire history has been altered, deleted, or
added. Nor do the changes typically affect records only at the beginning or only at the end of a
history. It is in this sense that alterations, deletions, and additions look selective, which in turn will
affect the trading signals an analyst generates and hence her track record.
A majority of anonymizations represent the removal of entire histories: 11,642 (8,670+2,972)
cases where an analyst’s entire history for a ticker has been anonymized (including cases where the
analyst covered a ticker only once), and 1,210 cases where every recommendation by an analyst has
been anonymized. In addition, 31.7% of the 19,904 anonymized recommendations affect selective
parts of an analyst’s history, including 1,017 cases where names are missing intermittently.
III.B Track Records of Analysts: Portfolio Returns to Anonymized Recommendations
In this section, we explore the consequences of these distortions to analysts’ track records.
21
Specifically, we examine the stock return performance of anonymized and non-anonymized
recommendations, for the sample of analysts for whom some but not all of their past
recommendations are anonymized. We focus our analysis here on the anonymizations, since our
pattern analysis above indicates that the anonymizations are associated with particularly large
changes to a given analyst’s historical track record; but we also discuss related tests for the other
three types of changes below. We formally test the hypothesis that anonymized recommendations
perform differently from non-anonymized recommendations by computing calendar-time buy-and-
hold portfolio returns as in Barber, Lehavy, and Trueman (2005) and Barber et al. (2006). For each
type of recommendation (anonymized or non-anonymized), we form two portfolios: (1) An
“upgrade” portfolio, consisting of all stocks that at least one analyst upgraded to buy or strong buy
from a previous hold, sell, or strong sell recommendation, or initiated coverage on with a buy or
strong buy rating; and (2) a “downgrade” portfolio, comprised of all stocks that at least one analyst
downgraded to hold, sell, or strong sell from a previous buy or strong buy recommendation, or
initiated coverage on with a hold, sell, or strong sell rating. We include holds in the downgrade
portfolio since the distribution of recommendations suggests that many analysts effectively rated
stocks on a three-point scale (hold/buy/strong buy).
The construction of these portfolios closely follows the methodology in Barber, Lehavy, and
Trueman (2005) and Barber et al. (2006). For each recommendation that is eligible for the upgrade
portfolio, for example, the recommended stock enters the portfolio at the close of trading on the day
the recommendation is announced. This approach explicitly excludes the announcement day return,
on the assumption that many investors likely become aware of recommendation changes only with a
delay. Each recommended stock remains in the portfolio until either the stock is downgraded or
dropped from coverage by the analyst, for up to one calendar year (after which time the stock is
automatically removed from the portfolio). If more than one analyst recommends a particular stock
22
on a given date, the stock will appear multiple times in the portfolio on that date (once for each
recommendation).
Assuming an equal dollar investment in each recommendation, the portfolio return on date t is
given by ∑∑ ==
tt n
i it
n
i itit xxR11
/ , where Rit is the date t return on recommendation i, nt is the number of
recommendations in the portfolio, and xit is the compounded daily return of recommended stock i
from the close of trading on the day of the recommendation through day t-1. (The variable xit equals
one for a stock recommended on day t-1.) The portfolio is updated daily, so that stocks that are
downgraded or dropped from coverage are taken out of the portfolio at the close of trading on the
day of the downgrade/drop. Calendar-time daily returns for the downgrade portfolio are computed
in the same manner. We calculate abnormal portfolio returns exactly as in Section II.D, with the
exception that the analysis here employs daily returns instead of monthly returns.23
Panel A of Table IX presents the average daily abnormal returns to the upgrade portfolios of
anonymized and non-anonymized recommendations. Over the entire 1993-2002 sample period,
anonymized recommendations significantly underperform non-anonymized recommendations. The
abnormal return on a long-short portfolio that goes long non-anonymized recommendations and
short anonymized recommendations is statistically significant and economically important, earning
alphas of between 1.4 and 1.5 basis points per day (3.6% and 3.8% annualized).
We next divide the sample into two sub-periods: (1) The period of the late 1990s bull market
(i.e., the period ending March 10, 2000, the date of the NASDAQ market peak), and (2) the period
of the bear market (i.e., the period beginning on March 10, 2000 and extending to the end of our
sample). This division reveals an interesting picture, and one that complements recent evidence that
analysts were issuing overly optimistic recommendations and forecasts with little regard to
23 Factors are obtained from Kenneth French’s website. In addition to the factors listed in eq. (1), we include one lag of
each independent variable in all regressions to control for nonsynchronous trading (Scholes and Williams (1977)).
23
fundamentals at the height of the dot-com frenzy (see Hong and Kubik (2003) and Barber, Lehavy,
and Trueman (2005)). As columns (5) and (6) of Panel A show, the poor performance of
anonymized upgrades-to-buy is particularly pronounced in the post-bubble period, with daily risk-
adjusted underperformance ranging from an economically large 2.4 to 2.7 basis points (6.1% to
6.8% annualized). Meanwhile, underperformance by non-anonymized upgrades-to-buy in this
period is minimal, and statistically insignificant. Abnormal returns on long-short portfolios designed
to capture this differential performance between anonymized and non-anonymized upgrades-to-buy
are again statistically significant and economically large: Up to 6.6% annualized.
We next refine our tests by focusing on analysts who remain active in the profession after 2002.
Panel B of Table IX reports results for this sub-sample of analysts and indicates that, once again,
anonymized upgrades-to-buy significantly underperform their benchmarks and their non-
anonymized counterparts.24 The corresponding long-short portfolio earns between 1.5 and 1.6 basis
points in daily abnormal returns over the full sample period (3.8% to 4.0% annualized), and
between 2.7 and 2.9 basis points in the bear market period (6.7% to 7.4% annualized); the t-
statistics range between 2.36 and 3.10.
Since our results in Table IV indicate that bolder recommendations are more likely to be
anonymized, we also examine the return performance of a sub-sample of “bold” upgrades-to-buy.
Specifically, we restrict the sample to recommendations that deviate in absolute terms from the
average recommendation for the stock that year by at least one notch (e.g., if the consensus is a
hold, a buy recommendation by an analyst would be flagged as bold). Panel C of Table IX shows
that anonymized upgrades that are bolder in nature are particularly poor performers, with average
underperformance relative to the benchmarks as high as 2.6 basis points in the full sample period
24 Screening out analysts with very few or very many anonymizations yields similar results, as does “un-restricting” the
sample by comparing the full sample of anonymizations to the full sample of non-anonymized recommendations.
24
(6.7% annualized) and 4.7 basis points in the bear market period (11.9% annualized). Non-
anonymized bold upgrades also underperform their benchmarks, but not nearly as dramatically; as a
result, the corresponding long-short portfolio earns up to 2.1 basis points in the full sample period
(5.2% annualized), and 3.6 basis points in the bear market period (9.1% annualized).
Panel D presents a different picture. In the full sample, and within each sub-period, a portfolio
of anonymized downgrades to hold/sell does not perform in a statistically or economically different
manner from a portfolio of non-anonymized downgrades.25
In unreported tests we also replicate these tests for the deletions, additions, and alterations. We
find that altered upgrades-to-buy perform significantly worse than non-altered upgrades-to-buy, but
the economic significance of this result is modest. We also find some evidence that additions
perform better than non-additions (particularly for upgrades in the pre-bubble period), but these
results are only marginally statistically significant. Lastly, we find no evidence that the performance
of deleted recommendations differs significantly from non-deleted recommendations.
Interestingly, searching over a wide variety of specifications, we could not identify a single case
where any of the four types of changes was economically or statistically significant in the direction
that would worsen the track records of the affected analysts. All of these findings could prove
problematic for studies of stock-picking skill persistence (see Mikhail, Walther, and Willis (2004)).
III.C Career Outcomes of Analysts
The evidence in Tables IV and IX suggests that it is the boldest, worst performing
recommendations that were most prone to be anonymized. As a consequence, the track records of
the analysts concerned now look better than they should. In this section, we extend this analysis to
study analyst job movements and show that the growing body of academic research into the career
25 Results (not shown) for the sub-sample that excludes potential retirees and for the sub-sample of bold hold
recommendations are similar.
25
concerns of analysts will be affected by the changes we document in the I/B/E/S historical
recommendations database.
Our analysis closely follows previous work by Hong, Kubik, and Solomon (2000) and Hong and
Kubik (2003) which relates career outcomes (such as changes of employer, promotions, demotions,
and exit from the profession) to a total of five measures of analyst ability (forecast accuracy,
forecast optimism, forecast boldness, all-star ranking, and seniority).26 We add to this list of
explanatory variables four indicators identifying analysts whose recommendations have partially
been altered, dropped, added to, or anonymized.27 To the extent that these changes (and in
particular, the anonymizations) make an analyst’s track record less transparent, we expect some
analysts to experience more favorable career outcomes than their characteristics and abilities would
ordinarily warrant.
We focus on the group of U.S. based analysts who contributed earnings forecasts to I/B/E/S in
2002 (that is, the individuals whose recommendations are most likely to have changed between the
2002 and 2004 tapes) and follow their career progressions through December 2005. This results in
7,696 analyst-years. Like Hong, Kubik, and Solomon (2000) and Hong and Kubik (2003), we do
not observe whether an analyst is fired, demoted, or promoted. We can, however, observe breaks in
an analyst’s contributions to I/B/E/S (which Hong, Kubik, and Solomon interpret as an adverse
career outcome) as well as moves to another employer. We follow Hong, Kubik, and Solomon in
calling a promotion a move from a small (fewer than 25 analysts) to a large (25 or more analysts)
26 Hong, Kubik, and Solomon’s (2000) relative forecast accuracy is defined as in Section I, while their relative forecast
boldness is the scaled rank of the analyst’s average absolute deviation from the earnings forecast consensus. It is
constructed analogously to relative forecast accuracy. For Hong and Kubik’s (2003) relative forecast optimism, we set a
dummy variable to one if an analyst’s forecast for a stock between January 1 and July 1 of year t exceeds the consensus
forecast; and zero otherwise. The analyst’s relative forecast optimism is then defined as the average of the dummy
variables over all stocks followed by the analyst in years t-2 through t. 27 We conservatively impose a 5% filter on the anonymization indicator variable, such that it equals one if 5% or more
of the analyst’s recommendations were anonymized. Our results are not sensitive to this coding. They are also not
sensitive to the inclusion of anonymizations dating back to the 1990s.
26
brokerage firm, and vice versa for a demotion.28 We also track whether an analyst is rated the top
stick picker in her sector in the following spring’s Wall Street Journal “Best on the Street” awards,
which are exclusively based on recommendations data supplied by Thomson Financial and vetted
by the analysts and brokerage firms.29
We code an analyst as leaving the profession if she contributed forecasts to I/B/E/S in year t but
not in year t+1. (Our coding allows the analyst to return in, say, year t+2 and so identifies periods of
unemployment as temporary exits from the profession.) We similarly track job moves by comparing
the identity of the analyst’s employer at the end of years t and t+1. The dataset thus consists of an
unbalanced annual panel tracking analysts across brokerage firms over the years 2003 to 2005. To
ensure we are tracking the right individuals, we correct thousands of errors in the I/B/E/S Broker
Translation (“bran”) file primarily using the NASD BrokerCheck service and Nelson’s Directory of
Investment Research. The most common error involves an analyst being assigned a new identifier
after she has moved to a new employer.30 Without these corrections, the number of exits from the
industry would be vastly overstated while the number of job moves would be understated. We are
also careful to take mergers among brokerage firms into account. Mergers result in analysts at the
target being assigned to the acquirer’s I/B/E/S broker code after the merger, giving the false
appearance of a job move.
The aforementioned five measures of analyst ability that Hong, Kubik, and Solomon (2000) and
28 Hong and Kubik report that results are generally robust to other definitions of “high status” employers (including the
number of all-star analysts employed and the reputation of the brokerage firm in the IPO market). 29 The Wall Street Journal reports, “The data on which the rankings are based were subjected to a rigorous verification
process […]. For example, Thomson Financial created a special Web site that allowed analysts to see the underlying
data on which they were to be evaluated and request changes in any information they considered inaccurate.” (5/12/03) 30 We make four types of corrections: (1) Around 3,000 cases where an analyst is assigned multiple identifiers, often
following a job move or name change; (2) more than 200 cases where two people are assigned the same identifier; (3)
nearly 1,000 analysts who have data credited under an incorrect broker code, creating the impression that they moved
back and forth between employers; and (4) around 60 cases where the wrong analyst is credited with a recommendation.
We also ensure that analyst identifiers are consistent between the recommendations and earnings forecast tapes.
Because our merge of the 2002 and 2004 tapes does not condition on analyst identifiers, none of these data clean-ups
cause the changes we document, though they allow us to measure analyst characteristics and job moves cleanly.
27
Hong and Kubik (2003) explore are practically orthogonal to each other, with two exceptions:
Relatively more accurate analysts tend to be bolder (correlation: 23.5%) and all-stars tend to have
been in the industry for longer (correlation: 19.2%). We thus depart from Hong, Kubik, and
Solomon by simultaneously including all five measures in our empirical specifications rather than
including them one by one. We do so to conserve space; our results are not sensitive to this choice.
We estimate multivariate probits that in addition control for year fixed effects, the log number of
stocks the analyst covered in year t, and the log number of analysts who covered the analyst’s
average stock. The latter two variables are included based on Hong, Kubik, and Solomon’s
argument that analysts who cover few companies, and who face less competition from other
analysts, are more likely to score in the tails of the distribution of accuracy, boldness, and optimism.
Table X reports the results. The dependent variable in the first column is an indicator variable
equal to one if the analyst changed employers between t and t+1. The unconditional probability of a
job move is 11.5% in our sample. At the means of the other covariates, this increases by a
statistically and economically significant 26.6% (to 14.5%) for analysts whose track record through
July 2002 has been partly anonymized, and by 16.9% for analysts with data additions. Neither
deletions nor alterations affect the likelihood of a job move significantly. Analysts who are more
senior, who are bolder relative to their peers in the same sector, who cover more stocks, and whose
stocks are covered by fewer rival analysts are significantly less likely to move jobs. The predicted
probability of an all-star moving employers in our time period is close to zero (2.2% versus 11.5%
for the sample as a whole). The pseudo-R2 of 3.8% indicates that a substantial portion of the
variation in outcomes remains unexplained, though the R2 rises to 11.3% when we include random
brokerage firm effects (not tabulated).31
31 Hong, Kubik, and Solomon (2000) and Hong and Kubik (2003) include a full set of brokerage firm fixed effects. As
fixed effects probit results in biased coefficient estimates, we have instead investigated random-effects panel probits.
Our results are virtually identical statistically and economically whether or not we include random effects.
28
We conclude that analysts with anonymizations and additions are more likely to move jobs, but
a job move could be good or bad news depending on whether it amounts to a promotion or a
demotion. In the next two columns we focus on moves to a large brokerage firm as a proxy for a
favorable career outcome. Controlling for the fact that analysts are less likely to move to a large
brokerage firm if they most recently worked for a small brokerage firm (p=0.057), the specification
in column (2) shows that the probability of being hired by a large brokerage firm is positively and
significantly related to both anonymizations and additions (p<0.001). Economically, anonymization
status is the second most significant determinant in this specification (after all-stars, who we saw
earlier hardly move anywhere during our sample period): It increases the likelihood of moving to a
large brokerage firm by 2.8 percentage points, an increase of 53.5% from the unconditional
likelihood of 5.2%.32 Additions increase that likelihood by 36.1%. More senior analysts, those who
make relatively bolder forecasts, and those covering under-researched stocks are significantly less
likely to move to a large brokerage firm.
In column (3), we restrict the sample to analysts who work for a small brokerage firm in year t
and ask whether anonymization helps them move up to a large brokerage firm in t+1. The relevant
sample size here is 2,468 analyst-years. Because all-star status perfectly predicts such a move
(every “promotion” involves an all-star) we cannot include it in this specification. If anything,
anonymization status now has an even greater effect than before. Analysts whose names were
removed from some I/B/E/S records see their chances of moving up to a large brokerage firm
improve by 63.5%, from 4.7% to 7.7% (p=0.006). This effect is easily larger than any other in this
model, followed by additions, which improve the likelihood by 47.6%.
Columns (4) and (5) repeat the analysis shown in columns (2) and (3) for moves to small
32 In unreported tests, we find that the probability of moving to a large brokerage firm peaks when 20 to 30 of an
analyst’s records have been anonymized.
29
brokerage firms. The coefficients estimated for anonymization and addition status are slightly
positive but both statistically and economically insignificant in either model. Thus, while a changed
track record is associated with promotions, it is unrelated to demotions. Interestingly, analysts
whose recommendations were altered are between 20% and 30% more likely to be demoted (with
p=0.05 in column (4) and p=0.023 in column (5)). Among the controls, we find that all-stars
virtually never move to a small brokerage firm whereas more optimistic analysts and those who
cover fewer stocks that are more widely covered are more likely to end up at a smaller house.
In column (6), we investigate temporary or permanent departures from the industry. Analysts
are significantly more likely to suffer unemployment or retire from Wall Street if they are not all-
stars, or top stock pickers, have been in the industry for longer, or have a reputation for relatively
inaccurate research, consistent with the findings reported in Hong, Kubik, and Solomon (2000) and
Hong and Kubik (2003). On the other hand, relatively more productive analysts, measured by the
number of companies covered, are more likely to remain in employment, especially if there is more
competition from other analysts. Controlling for these factors, we find that changes to an analyst’s
historic recommendations have no effect on the likelihood of leaving the profession.
The final model, reported in column (7), shows that an analyst is more likely to be rated the top
stock picker in her sector if some of her recommendations have been dropped (p=0.043) or added to
(p=0.048), increasing her chances by 47.8% and 40.4%, respectively. The effect of anonymizations
is large as well, at 41%, but only marginally statistically significant (p=0.067).33 Beyond these, there
are only two significant determinants of WSJ status: Top stock pickers cover more stocks, and cover
stocks that fewer rivals cover, though these effects are somewhat smaller economically.
33 However, the p-value becomes 0.047 if we define anonymizers using only recent (2000-2002) anonymizations.
30
IV. Conclusions
Our main contribution is to alert researchers to widespread, and previously undocumented, ex
post changes to an important financial database. Comparing two snapshots of the entire I/B/E/S
analyst stock recommendations database, taken in 2002 and 2004 but each covering the same time
period 1993-2002, we identify tens of thousands of changes. We isolate four particular types of
changes which we term alterations, deletions, additions, and anonymizations. Collectively, we
identify 54,729 changes to a database originally containing 280,463 observations. The changes
appear to affect brokerage firms, analysts, and tickers in a non-random manner.
We demonstrate that the changes have a large and significant effect on several features of the
data that are routinely used by academics and practitioners, including: 1) The distribution of
recommendations across all stocks and for individual stocks; 2) the bullishness of recommendations
at the level of individual brokerage firms; 3) the classification of trading signals such as upgrades
and downgrades; 4) consensus recommendations in individual stocks; and 5) the performance track
records of individual analysts. We further illustrate the significance of these changes by showing
how they affect back-tests of a popular trading strategy and analysts’ career outcomes in the three
years since the changes occurred.
Much like problems in other financial databases such as CRSP and Compustat have been shown
to cause biases in academic research, our findings have important ramifications for existing and
future empirical studies of equity analysts.
31
References
Asker, John, and Alexander Ljungqvist, 2005, Sharing investment bankers, mimeo, NYU.
Barber, Brad, Reuven Lehavy, and Brett Trueman, 2005, Comparing the stock recommendation
performance of investment banks and independent research firms, Journal of Financial
Economics, forthcoming.
Barber, Brad, Reuven Lehavy, Maureen McNichols, and Brett Trueman, 2001, Can investors profit
from the prophets? Security analyst recommendations and stock returns, Journal of Finance
56, 531-564.
Barber, Brad, Reuven Lehavy, Maureen McNichols, and Brett Trueman, 2003, Prophets and losses:
Reassessing the returns to analysts’ stock recommendations, Financial Analysts Journal 59,
88-96.
Barber, Brad, Reuven Lehavy, Maureen McNichols, and Brett Trueman, 2006, Buys, holds, and
sells: The distribution of investment banks’ stock ratings and the implications for the
profitability of analysts’ recommendations, Journal of Accounting and Economics,
forthcoming.
Bennin, Robert, 1980, Error rates in CRSP and COMPUSTAT: A second look, Journal of Finance
35, 1267-1271.
Canina, Linda, Roni Michaely, Richard Thaler, and Kent Womack, 1998, Caveat compounder: A
warning about using the daily CRSP equal-weighted index to compute long-run excess
returns, Journal of Finance 53, 403-416.
Carhart, Mark, 1997, On persistence in mutual fund performance, Journal of Finance 52, 57-82.
Cowen, Amanda, Boris Groysberg, and Paul M. Healy, 2003, Which types of analyst firms make
more optimistic forecasts? Mimeo, Harvard Business School.
Elton, Edwin J., Martin J. Gruber, and Christopher R. Blake, 2001, A first look at the accuracy of
the CRSP Mutual Fund Database and a comparison of the CRSP and Morningstar Mutual
Fund Databases, Journal of Finance 56, 2415-2430.
Francis, Jennifer, and Donna Philbrick, 1993, Analysts’ decisions as products of multi-task
environment, Journal of Accounting Research 31, 216-230.
Hong, Harrison, and Jeffrey D. Kubik, 2003, Analyzing the analysts: Career concerns and biased
forecasts, Journal of Finance 58, 313-351.
Hong, Harrison, Jeffrey D. Kubik, and Amit Solomon, 2000, Security analysts’ career concerns and
herding of earnings forecasts, RAND Journal of Economics 31, 121-144.
Irvine, Paul, 2004, Analysts’ forecasts and brokerage-firm trading, Accounting Review 79, 125-150.
Jackson, Andrew, 2005, Trade generation, reputation, and sell-side analysts, Journal of Finance 60,
32
673-717.
Jegadeesh, Narasimhan, Joonghyuk Kim, Susan D. Krische, and Charles Lee, 2004, Analyzing the
analysts: When do recommendations add value?, Journal of Finance 59, 1083-1124.
Kadan, Ohad, Leonardo Madureira, Rong Wang, and Tzachi Zach, 2005, Conflicts of interest and
stock recommendations, mimeo, Washington University.
Kothari, S.P., 2001, Capital markets research in accounting, Journal of Accounting and Economics
31, 105-231.
Lim, Terence, 2001, Rationality and analysts’ forecast bias, Journal of Finance 56, 369-385.
Lin, Hsiou-wei, and Maureen F. McNichols, 1998, Underwriting relationships, analysts’ earnings
forecasts and investment recommendations, Journal of Accounting and Economics 25, 101-
127.
Ljungqvist, Alexander, Felicia Marston, Laura Starks, Kelsey Wei, and Hong Yan, 2006, Conflicts
of interest in sell-side research and the moderating role of institutional investors, Journal of
Financial Economics, forthcoming.
Michaely, Roni, and Kent L. Womack, 1999, Conflict of interest and the credibility of underwriter
analyst recommendations, Review of Financial Studies 12, 653-686.
Mikhail, Michael B., Beverly R. Walther, and Richard H. Willis, 1999, Does forecast accuracy
matter to security analysts? Accounting Review 74, 185-200.
Mikhail, Michael B., Beverly R. Walther, and Richard H. Willis, 2004, Do security analysts exhibit
persistent differences in stock picking ability? Journal of Financial Economics 74, 6-91.
Ramnath, Sundaresh, Steve, Rock, and Philip Shane, 2005, A review of research related to financial
analysts’ forecasts and stock recommendations, mimeo, Georgetown University.
Rosenberg, Barr and Michel Houglet, 1974, Error rates in CRSP and Compustat data bases and their
implications, Journal of Finance 29, 1303-1310.
Scholes, Myron and Joseph Williams, 1977, Estimating betas from nonsynchronous data, Journal of
Financial Economics 5, 309-327.
Shumway, Tyler, 1997, The delisting bias in CRSP data, Journal of Finance 52, 327-340.
Shumway, Tyler and Vincent A. Warther, 1999, The delisting bias in CRSP’s Nasdaq data and its
implications for interpretation of the size effect, Journal of Finance 54, 2361-2379.
Stickel, Scott E., 1992, Reputation and performance among security analysts, Journal of Finance
47, 1811-1836.
Womack, Kent, 1996, Do brokerage analysts’ recommendations have investment value? Journal of
Finance 51, 137-167.
33
Appendix A.
This appendix demonstrates the difficulty of “undoing” the changes we document in the paper.
A researcher using today’s version of the I/B/E/S historical recommendations database (downloaded
on February 12, 2007) could obviously not undo the deletions or additions without access to a
historic snapshot of the data, such as our 2002 tape. In theory, there are ways to potentially undo the
alterations, but such fixes are incomplete and will likely introduce errors into the data.
A. Anonymizations
In response to this paper, Thomson Financial have recently reinstated the missing analyst names
in the recommendation history file; as of February 12, 2007, the data on WRDS now reflect this
reinstatement.
Thomson Financial have stated that users of downloads/snapshots of the recommendations
database from the period May 2004 to December 2006 can repopulate the missing analyst names
from the I/B/E/S earnings forecast database. The feasibility of this procedure requires: 1) Consistent
coverage across the databases; 2) consistent analyst codes; and 3) that the relevant observations are
not also anonymized on the earnings tape. When we simulate this procedure using data from a
November 13, 2006 download, we find that only 48.7% of the anonymized recommendations have
an earnings forecast for the same standardized broker name, ticker, and date, 4.2% of which are
anonymous forecasts or incorrect matches compared to the historical 2002 tape. Widening the
match window results in further matches but also increases the error rate. Within a 180-day (30-
day) window before and after the recommendation date, an earnings forecast for the same
standardized broker name and ticker can be found for 80.1% (52.5%) of the anonymized
recommendations, 5.8% (3.2%) of which are anonymous forecasts or incorrect matches. Thus,
around a quarter of anonymized recommendations cannot be reliably identified using the I/B/E/S
earnings forecast database.
34
B. Alterations
As noted in the text, the broker recommendation text field does not always change when an
I/B/E/S recommendation level is altered. In these cases, it may be possible to use the broker text to
recreate the original 2002 recommendation. However, there is an inherent difficulty with this
approach in that there is no one-to-one mapping between broker text categories and I/B/E/S
recommendation levels even among non-altered recommendations. (Approximately one third of
brokers without altered records map more than one broker text category to a single I/B/E/S
recommendation level.)
Ignoring this practical difficulty, by virtue of the fact that we know which recommendations
were actually altered, we find that 25.7% of the alterations are accompanied by broker text changes.
C. Deletions
The text documents the removal of 4,923 historical recommendations as of 2004. Surprisingly,
by the end of 2006, approximately one-half of these deletions have been reversed and thus re-
appear in the database (as of November 13, 2006). On the other hand, there are an additional 6,268
deletions of records from the 1993-2002 period that occurred between 2004 and 2005. Most of these
(93%) result from the removal of five brokers from the database.
The frequent adding and deleting of historical records imply that the recommendations database
has been unstable over time, making comparisons across time difficult.
35
Figure 1. Changes to Trading Signals.
Original upgrades
0%
5%
10%
15%
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
All trading signals on the 2002 tape
0%
5%
10%
15%
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
Original downgrades
0%
1%
2%
3%
4%
5%
6%
7%
8%
9%
10%
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
All trading signals on the 2002 tape ("brkcd"-level)
0%
5%
10%
15%
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
Original reiterations
0%
5%
10%
15%
20%
25%
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
The 2002 I/B/E/S tape contains 280,463 investm
ent recommendations issued between O
ctober 1993 and July 18, 2002. We match up the 2002 and 2004 tapes by
standardized broker name, I/B/E/S ticker, and recommendation date. O
bservations on the 2004 tape dated after July 18, 2002 are ignored. Trading signals are
constructed on a per-broker and per-I/B/E/S-ticker basis using a twelve-m
onth look-back w
indow. For instance, an upgrade (downgrade) is defined as a positive
(negative) change from a recommendation issued by the same broker for the same I/B/E/S ticker w
ithin the previous twelve months. R
eiterations are repeated
recommendations at the same recommendation level. If the previous recommendation w
as issued m
ore than twelve months ago, the current recommendation is
defined to be a reinitiation. If there is no previous recommendation, the current recommendation is defined to be an initiation. The graphs show the annual fraction
of trading signals that are classified differently if the 2004 tape is used instead of the 2002 tape. Initiations and reinitiations are included in the two “All trading
signals…
” graphs, w
hich differ from each other depending on w
hether recommendations are arranged by our standardized broker name, or the I/B/E/S brokerage
firm
code (“brkcd”).
36
Table I. Summary Statistics for Alterations, Deletions, Additions, and Anonymizations. The 2002 I/B/E/S tape contains 280,463 investment recommendations issued between October 1993 and July 18, 2002. We
match up the 2002 and 2004 tapes by standardized broker name, I/B/E/S ticker, and recommendation date. Observations on
the 2004 tape dated after July 18, 2002 are ignored. We define an alteration as a record that has a different recommendation
level according to the two tapes (i.e., a change from strong buy, buy, hold, underperform, sell); a deletion as a record that
appears on the 2002 tape but not on the 2004 tape; an addition as a record that appears on the 2004 tape but not on the 2002
tape; and an anonymization as a record attributed to a specific analyst on the 2002 tape whose name is missing on the 2004
tape (and whose amaskcd identifier has been set to zero). We exclude from our analysis 22,240 additions by brokerage
firms that did not contribute data to I/B/E/S according to the 2002 tape, appeared on the 2004 tape, and have since been
dropped from I/B/E/S again. On closer inspection, these are quantitative research shops which produce algorithmic
recommendations that are constrained to be symmetrically distributed.
Panel A: Breakdown of Types of Change By Year
Entire
2002
sample Alterations Deletions
Altered
or
deleted Additions
Anonymizations
Total
changes
ex post
No. No. % No. % % No. % No. % No.
Total 280,463 10,698 3.8% 4,923 1.8% 5.6% 19,204 6.8% 19,904 7.1% 54,729
1993 14,428 754 5.2% 91 0.6% 5.9% 1,459 10.1% 1,045 7.2% 3,349
1994 28,204 961 3.4% 313 1.1% 4.5% 2,773 9.8% 2,006 7.1% 6,053
1995 29,364 954 3.2% 330 1.1% 4.4% 2,206 7.5% 2,276 7.8% 5,766
1996 28,260 1,027 3.6% 365 1.3% 4.9% 1,946 6.9% 2,118 7.5% 5,456
1997 28,885 933 3.2% 417 1.4% 4.7% 1,454 5.0% 2,161 7.5% 4,965
1998 33,453 1,302 3.9% 682 2.0% 5.9% 1,513 4.5% 2,335 7.0% 5,832
1999 34,879 1,174 3.4% 748 2.1% 5.5% 2,572 7.4% 2,435 7.0% 6,929
2000 30,868 1,248 4.0% 602 2.0% 6.0% 2,147 7.0% 2,395 7.8% 6,392
2001 30,339 1,337 4.4% 511 1.7% 6.1% 2,557 8.4% 1,928 6.4% 6,333
2002 21,783 1,008 4.6% 864 4.0% 8.6% 577 2.6% 1,205 5.5% 3,654
Panel B: Breakdown of Distribution of Recommendations by Type of Change
Alterations
2002 tape Pre-alteration Post-alteration Deletions Additions Anonymizations
Rec. level No. % No. % No. % No. % No. % No. %
All 280,463 10,698 10,698 4,923 19,204 19,904
1 (strong buy) 80,260 28.6 3,314 31.0 4,432 41.4 2,586 52.5 3,923 20.4 5,896 29.6
2 102,904 36.7 5,302 49.6 3,413 31.9 1,363 27.7 4,738 24.7 7,204 36.2
3 88,199 31.4 1,486 13.9 1,205 11.3 883 17.9 9,166 47.7 6,108 30.7
4 4,644 1.7 136 1.3 1,093 10.2 28 0.6 951 5.0 367 1.8
5 (strong sell) 4,456 1.6 460 4.3 555 5.2 63 1.3 426 2.2 329 1.7
Mean 2.11 1.98 2.06 1.70 2.44 2.10
37
Panel C: Breakdown of Alterations by Direction of Alteration
No. of
alterations
Mean
alteration
Distribution of alterations
(negative = downgrade)
Ratio:
downgrade/
upgrade
-2 -1 1 2 3
Total 10,698 -0.075 749 4,630 5,311 7 1 1.01
1993 754 -0.050 114 225 415 0.82
1994 961 -0.101 118 352 491 0.96
1995 954 0.149 86 277 591 0.61
1996 1,027 -0.046 94 396 537 0.91
1997 933 -0.076 90 367 476 0.96
1998 1,302 -0.131 96 592 614 1.12
1999 1,174 0.068 69 444 660 1 0.78
2000 1,248 -0.119 53 619 575 1 1.17
2001 1,337 -0.108 17 716 603 1 1.21
2002 1,008 -0.304 12 642 349 4 1 1.85
38
Table II. Summary Statistics: Brokerage Firm Characteristics.
The 2002 I/B/E/S tape contains 280,463 investm
ent recommendations issued between October 1993 and July 18, 2002 by 7,817 separate named analysts working for 385
distinct brokerage firm
s. For each of the four types of changes to the 1993-2002 recommendations data, w
e split the 385 brokerage firm
s into two groups depending on
whether any of their recommendations have been altered, deleted, added to, or anonymized. (Sixteen firms already only had anonymous recommendations on the 2002
tape, leaving 369 brokerage firm
s whose recommendations could be subject to anonymization.) Equity underwriting m
arket share is used as a proxy for the size of the
firm
’s investm
ent banking operations and is based on data from Thomson Financial/SDC’s U
.S. New Issues database. The number of analysts at each brokerage firm is
used as a proxy for the size of the firm
’s research departm
ent and is based on the number of separate named analysts who contribute recommendations to I/B/E/S in 2002.
A brokerage firm
is deemed to have exited sell-side research if by 2002 it no longer contributes data to I/B/E/S. Sanctioned brokerage firm
s are the 12 firms that w
ere
subject to the 2003 Global Settlement.
Alterations?
Deletions?
Additions?
Anonymizations?
Yes
No
Yes
No
Yes
No
Yes
No
Number of brokerage firm
s 64
321
118
267
130
255
239
130
Brokerage firm
size in 2002
mean equity underwriting m
arket share
1.3%
0.0%
0.8%
0.0%
0.7%
0.0%
0.4%
0.0%
mean # analysts at brokerage firm
42.3
5.6
32.9
2.3
21.4
1.3
18.4
0.8
Exited sell-side research by 2002?
Yes
1
104
7
98
11
94
0
89
No
63
217
111
169
119
161
239
41
Sanctioned in Global Settlement?
Yes
9
3
12
0
12
0
12
0
No
55
318
106
267
118
255
227
130
Investment banking operations in 2002?
Yes
29
37
48
18
57
9
61
4
No
35
284
70
249
73
246
178
126
39
Table III. Summary Statistics: Analyst Characteristics.
The 2002 I/B/E/S tape contains 280,463 investm
ent recommendations issued between October 1993 and July 18, 2002 by 7,817 separate named analysts working for 385
distinct brokerage firm
s. For each of the four types of changes to the 1993-2002 recommendations data, we split the 7,817 analysts into two groups depending on whether
any of their recommendations have been altered, deleted, added to, or anonymized. The number of analysts at each brokerage firm
is used as a proxy for the size of the
firm
’s research departm
ent and is based on the number of separate named analysts who contribute recommendations to I/B/E/S in 2002. Equity underwriting m
arket share
is used as a proxy for the size of the firm
’s investm
ent banking operations and is based on data from Thomson Financial/SDC’s U
.S. New Issues database. Sanctioned
brokerage firm
s are the 12 firms that w
ere subject to the 2003 G
lobal Settlem
ent. A
nalyst characteristics are reported as of 2002 or the date of the analyst’s retirement
(exit from I/B/E/S) if earlier. All-stars are analysts ranked as a top-three or runner-up analyst in their sector by Institutional Investor. A
Wall Street Journal top stock
picker is the first-ranked analyst in each industry. Relative forecast accuracy is a measure of the analyst’s average forecast accuracy across the stocks she has covered in
the three years to July 2002 (or in the three years to the date of her retirement, if she has left the industry before July 2002) relative to the other analysts covering the same
stocks. It is constructed as in H
ong, Kubik, and Salomon (2000) and ranges from 0 to 1, with a higher number indicating greater forecast accuracy. As a proxy for
seniority, we compute the number of years since the analyst first appeared in the I/B/E/S database. A
nalysts are deemed to rem
ain in the profession post-2002 if they
contribute recommendations to I/B/E/S in 2003 through 2005. This m
ay undercount the number of analysts who remain in the profession. We use *
** and *
* to denote
significance at the 0.1% and 1% level, respectively, in t-tests of differences in m
eans or proportions, as appropriate. NM = “not meaningful”.
Alterations?
Deletions?
Additions?
Anonymizations?
Yes
No
t- test
Yes
No
t- test
Yes
No
t- test
Yes
No
t- test
Number of analysts
940
6,877
870
6,947
1,874
6,059
2,137
5,680
Panel A: Number of recommendations
Number of changed recommendations per analyst
Mean
10.9
0
NM
5.5
0
NM
7.7
0
NM
9.3
0
NM
Min
1
0
NM
1
0
NM
1
0
NM
1
0
NM
Median
5
0
NM
2
0
NM
3
0
NM
2
0
NM
Max
128
0
NM
103
0
NM
138
0
NM
311
0
NM
Fraction of analyst’s recommendations changed
Mean
31.7%
0
NM
17.6%
0
NM
40.2%
0
NM
29.5%
0
NM
Median
22.2%
0
NM
6.7%
0
NM
13.0%
0
NM
12.9%
0
NM
Panel B: Employer characteristics
Mean # analysts at brokerage firm
in 2002
70.0
98.6
***
85.1
96.4
***
58.9
68.9
***
96.7
94.5
-
Mean 2002 equity underwriting m
arket share
1.9%
3.8%
***
3.0%
3.7%
***
2.8%
3.9%
***
3.6%
3.6%
-
Fraction working at a sanctioned brokerage firm
23.6%
42.6%
***
34.6%
41.0%
***
40.1%
41.3%
-
39.7%
40.5%
-
40
Table III. Continued
.
Alterations?
Deletions?
Additions?
Anonymizations?
Yes
No
t- test
Yes
No
t- test
Yes
No
t- test
Yes
No
t- test
Panel C: Analyst characteristics
Fraction of analysts ranked all-stars
10.5%
15.3%
***
16.8%
14.4%
-
17.2%
13.7%
***
16.6%
14.0%
**
Fraction of analysts ranked top stock pickers
1.2%
0.5%
**
1.1%
0.5%
***
0.9%
0.4%
**
0.9%
0.4%
**
Mean relative forecast accuracy
47.9%
47.5%
-
48.0%
47.5%
-
48.9%
47.0%
***
48.8%
47.1%
**
Mean seniority (years in I/B/E/S database)
7.9
7.1
***
7.7
7.1
**
8.1
6.9
***
7.9
6.9
***
Mean # recommendations in I/B/E/S
56.4
31.9
***
64.3
31.2
***
52.4
28.8
***
54.0
27.7
***
Panel D: Post-2002 career paths
Fraction of analysts remaining in profession post-2002
69.9%
51.1%
***
75.6%
50.7%
***
58.8%
50.8%
***
70.8%
46.9%
***
Fraction of changed recommendations originally m
ade
by analysts who remain in the industry post-2002
65.5%
n.a.
NM
71.3%
n.a.
NM
49.1%
n.a.
NM
84.7%
n.a.
NM
41
Table IV. Characteristics of Altered, Dropped, Added, or Anonymized Recommendations. The unit of observation in this table is a recommendation, issued between October 1993 and July 18, 2002. The
dependent variable is an indicator variable that equals one (1) if the record has been altered (i.e., it has a different
recommendation level according to the two tapes); (2) if the record has been deleted (i.e., it appears on the 2002 tape
but not on the 2004 tape); (3) if the record is an addition (i.e., it appears on the 2004 tape but not on the 2002 tape); or
(4) if the record has been anonymized (i.e., if the name of the recommending analyst was disclosed on the 2002 I/B/E/S
tape but no longer appears on the 2004 I/B/E/S tape); and zero otherwise. Boldness, defined as the absolute difference
between the level of the recommendation and the consensus, is measured using the 2002 data (i.e., before alterations,
deletions, or additions); it is thus not defined for additions. Whether the brokerage firm has exited sell-side research by
2002 is a perfect predictor of anonymization, and hence is not included in column (4). Other explanatory variables are
defined in Tables II and III. All models are estimated using probit MLE and include random brokerage firm effects.
Intercepts and dummies identifying the year in which the recommendation was originally made are not shown. Standard
errors are shown in italics. Note that random-effects probit does not support a White or cluster adjustment for
heteroskedasticity. We use ***, **, and
* to denote significance at the 0.1%, 1%, and 5% level (two-sided), respectively.
Alter-
ations Deletions Additions
Anonymi-
zations
(1) (2) (3) (4)
Analyst characteristics
relative forecast accuracy -0.183 -0.200* 0.091 -0.487
***
0.129 0.087 0.065 0.059
=1 if analyst is ranked an Inst. Investor all-star in Oct 2001 -0.779*** -0.088
* -0.123
** 0.280
***
0.067 0.034 0.043 0.038
=1 if analyst is top stock picker in WSJ in June 2002 0.038 -0.039 0.314*** 0.156
**
0.099 0.099 0.067 0.054
seniority (log years in I/B/E/S) 0.092** 0.058
* 0.192
*** 0.049
**
0.035 0.024 0.021 0.019
ln(analyst’s no. of recommendations) 0.045* 0.001 -0.238
*** -0.104
***
0.021 0.016 0.012 0.012
=1 if analyst still active post-2002 -0.017 0.222*** 0.247
*** 0.674
***
0.045 0.033 0.023 0.021
Brokerage firm characteristics
investment banking size (% eq. underwriting mkt share, 2002) 0.003 -0.043 -0.028** -0.057
0.014 0.033 0.010 0.389
size of research department (log no. analysts at broker in 2002) 0.294*** 0.326
*** 0.181
*** 0.169
***
0.038 0.065 0.032 0.050
=1 if brokerage firm sanctioned in Global Settlement -1.209*** -0.481 -0.311
*** -0.079
0.183 0.408 0.055 0.074
=1 if brokerage firm exited sell-side research by 2002 -0.646* 0.583
* -0.092
***
0.292 0.232 0.257
Characteristics of company/recommendation
boldness of original recommendation 0.380*** -0.078
*** 0.041
**
0.024 0.021 0.013
=1 if company has IB relationship with brokerage firm -0.040 -0.062 -0.178* -0.023
0.085 0.051 0.079 0.081
ln(aggregate amount of capital company raised in prior 5 years) 0.004 -0.006 -0.020*** 0.001
0.005 0.004 0.003 0.003
Diagnostics
Pseudo R2 75.8 % 55.3 % 60.8 % 47.3 %
Wald test: all coefficients=0 (χ2) 1,113***
1,009*** 1,031
*** 1,510
***
Likelihood ratio test: all random effects = 0 (p-vale of χ2 test) <0.001 <0.001 <0.001 <0.001
Number of observations 258,587 258,587 283,227 258,587
42
Table V. Effect of Alterations, Additions, and Deletions on the Distribution of Recommendations for 1993-2002 H
istory.
The 2002 I/B/E/S tape contains 280,463 investm
ent recommendations issued between O
ctober 1993 and July 18, 2002. We match up the 2002 and 2004 tapes by
standardized broker nam
e, I/B/E/S ticker, and recommendation date. Observations on the 2004 tape dated after July 18, 2002 are ignored. We define an alteration as
a record that has a different recommendation level according to the tw
o tapes (i.e., a change from strong buy, buy, hold, underperform
, sell); a deletion as a record
that appears on the 2002 tape but not on the 2004 tape; and an addition as a record that appears on the 2004 tape but not on the 2002 tape. Panel A
provides a
breakdown of the distribution of recommendations by the strength of the rating (using the I/B/E/S five-point scale) for the 2002 tape, each of the three types of
changes, and the resulting 2004 tape. T
he mean recommendation level shown in the final row is computed as the frequency w
eighted recommendation. The
distribution of changes shown in the final column refers to the distribution of the net differences between the 2002 and the 2004 tapes. In Panel B w
e compute the
proportion of recommendations for a given I/B/E/S ticker in a given year that are buys (including strong buys), holds, or sells (including strong sells). There are
49,097 ticker-years on the 2002 tape, w
ith an average of 5.7 recommendations per ticker and year. A
ll entries in this table refer to the average company in a given
year. Results for the median company are sim
ilar. As the first block of Panel B shows, the recommendation distribution for the average ticker-year becomes more
conservative in the wake of the alterations, additions, and deletions to the 2002 tape: For the average ticker, the proportion of buys decreases by one percentage
point, while the proportion of holds and sells increases by 0.4 and 0.6 percentage points, respectively. In a given year, the m
ajority of tickers experience no change in
their recommendation distribution; see the m
iddle block of Panel B. The third block shows the change in the proportion of buys, holds, or sells conditional on each
such change. For instance, the proportion of sells increased by 11 percentage points among the 11.7% of tickers with changes in the proportion of sells in 1993.
Thus, by construction, the mean conditional changes in the proportion of buys, holds, and sells in the third block do not add up to zero. Panel C shows the effect of
the changes on the distribution of recommendations at the level of individual brokerage firms. W
e rank brokerage firm
s every quarter based on the fraction of their
end-of-quarter outstanding recommendations that are buys or strong buys and group them into quintiles, using data from either the 2002 tape or the 2004 tape. The
Panel traces the migrations into different quintiles when we sw
itch from the 2002 tape to the 2004 tape. For instance, 18.7% of the most bullish brokers on the 2002
tape migrate into another quintile on the 2004 tape. The unequal numbers of quintile constituents are caused by ties.
Panel A: Effect of Changes on O
verall Distribution
Distribution of 1993-2002
history on 2002 tape
Distribution of 1993-2002
history on 2004 tape
Distribution of net differences
between 2002 and 2004 tape
Rec. level
No.
%
No.
%
Net
difference
% of net
difference
All
280,463
294,744
14,281
1 (strong buy)
80,260
28.6%
82,715
28.1%
2,455
17.2%
2
102,904
36.7%
104,390
35.4%
1,486
10.4%
3
88,199
31.4%
96,201
32.6%
8,002
56.0%
4
4,644
1.7%
6,524
2.2%
1,880
13.2%
5 (strong sell)
4,456
1.6%
4,914
1.7%
458
3.2%
Mean
2.11
2.14
2.75
43
Panel B: Effect of Changes on Ticker-Level Distributions
Mean percentage point change in
proportion of
Fraction of I/B/E/S tickers with changes
in proportion of
Mean conditional percentage point change
in proportion (conditional on change in
proportion) of
No. tickers
buys
holds
sells
buys
holds
sells
buys
holds
sells
1993
3,629
-0.6
-0.7
1.3
26.7%
29.4%
11.7%
-2.3
-2.3
11.0
1994
4,610
-0.5
-0.3
0.7
29.1%
30.8%
14.1%
-1.6
-0.9
5.2
1995
4,903
-0.2
-0.4
0.6
26.8%
28.1%
11.2%
-0.9
-1.3
5.4
1996
5,430
-0.7
0.2
0.5
22.7%
24.2%
8.5%
-3.0
0.9
5.5
1997
5,627
-1.5
0.9
0.6
18.0%
19.7%
5.8%
-8.4
4.6
10.4
1998
5,773
-1.4
0.9
0.4
18.9%
20.5%
5.4%
-7.2
4.5
8.0
1999
5,623
-1.3
0.7
0.6
25.8%
27.4%
8.3%
-5.0
2.6
6.8
2000
5,114
-1.6
1.1
0.5
25.1%
26.6%
6.4%
-6.4
4.2
7.3
2001
4,481
-1.6
1.1
0.5
30.9%
31.5%
9.0%
-5.3
3.6
5.4
2002
3,907
-0.2
-0.1
0.3
19.1%
19.5%
6.9%
-1.0
-0.7
4.9
1993-2002
49,097
-1.0
0.4
0.6
24.1%
25.5%
8.5%
-4.1
1.6
6.8
Panel C: Migrations in Q
uarterly Broker-recommendation Q
uintiles (2002 tape vs. 2004 tape)
2002 tape
All changes
Quintile accord
ing to 2004 tape
No.
No.
%
Q1
Q2
Q3
Q4
Q5
Quintile as of 2002 tape
Q1 (most bullish)
1,644
307
18.7%
114
85
67
41
Q2
1,001
194
19.4%
85
44
45
20
Q3
1,309
290
22.2%
13
214
36
27
Q4
1,352
257
19.0%
10
23
181
43
Q5 (most conservative)
1,234
146
11.8%
6
13
25
102
All broker-quarters
6,540
1,194
18.3%
114
364
335
250
131
44
Table VI. Effect of Alterations, Additions, and Deletions on Trading Signals.
The 2002 I/B/E/S tape contains 280,463 investment recommendations issued between O
ctober 1993 and July 18, 2002. We match up the 2002 and 2004 tapes by
standardized broker name, I/B/E/S ticker, and recommendation date. O
bservations on the 2004 tape dated after July 18, 2002 are ignored. We define an alteration as a
record that has a different recommendation level according to the two tapes (i.e., a change from strong buy, buy, hold, underperform
, sell); a deletion as a record that
appears on the 2002 tape but not on the 2004 tape; and an addition as a record that appears on the 2004 tape but not on the 2002 tape. Panel A provides a breakdown of the
distribution of trading signals for the 2002 tape, each of the three types of changes, and the resulting 2004 tape. Trading signals are constructed on a per-broker and per-
I/B/E/S-ticker basis using a twelve-m
onth look-back w
indow. For instance, a downgrade is defined as a negative change from a recommendation issued by the same
broker for the same I/B/E/S ticker within the previous twelve months. If the previous recommendation was issued more than twelve months ago, the current
recommendation is defined to be a reinitiation. If there is no previous recommendation, the current recommendation is defined to be an initiation. Note that the net number
of changed trading signals is less than the sum of the trading signals that are changed due to alterations, additions, and deletions, as some changes cancel each other out.
The distribution of changes shown in the final column of Panel A refers to the distribution of the net differences between the 2002 and the 2004 tapes. Panel B provides a
transition m
atrix for the changed trading signals from 2002 to 2004.
Panel A: Effect of Changes on the Distribution of Trading Signals
Distribution of
Distribution of alterations
Distribution of
2002 tape
Pre-alteration
Post-alteration
deletions
additions
2004 tape
Trading signal
No.
%
No.
%
No.
%
No.
%
No.
%
No.
%
Net
changes
All signals
280,463
10,698
10,698
4,923
19,204
294,744
downgrade
56,464
20.1%
1,568
14.7%
1,063
9.9%
603
12.2%
5,215
27.2%
60,316
20.5%
27.0%
upgrade
46,843
16.7%
1,525
14.3%
580
5.4%
1,211
24.6%
2,312
12.0%
49,271
16.7%
17.0%
reiteration
21,950
7.8%
834
7.8%
2,284
21.4%
966
19.6%
1,794
9.3%
25,622
8.7%
25.7%
initiation
103,235
36.8%
4,673
43.7%
4,673
43.7%
1,357
27.6%
6,794
35.4%
105,774
35.9%
17.8%
reinitiation
51,971
18.5%
2,098
19.6%
2,098
19.6%
786
16.0%
3,089
16.1%
53,760
18.2%
12.5%
Panel B: Migrations in Trading Signals (2002 tape vs. 2004 tape)
2002 tape
All changes
Trading signal accord
ing to 2004 tape
Trading signal as of 2002 tape
No.
No.
%
downgrade
upgrade
reiteration
initiation
reinitiation
deleted
downgrade
56,464
2,981
5.3%
137
1,797
263
181
603
Upgrade
46,843
2,913
6.2%
68
1,581
11
42
1,211
reiteration
21,950
2,420
11.0%
457
935
22
40
966
initiation
103,235
4,811
4.7%
902
1,014
630
908
1,357
reinitiation
51,971
2,475
4.8%
365
942
271
111
786
added in 2004
19,204
5,215
2,312
1,794
6,794
3,089
all signals on 2002 tape
280,463
34,804
12.4%
7,007
5,340
6,073
7,201
4,260
4,923
45
Table VII. Effect of Alterations, Additions, and Deletions on Consensus Recommendations. This table reports descriptive statistics based on consensus analyst recommendations for the pre-2000 period. We use all
I/B/E/S recommendations that have been outstanding for less than one year. Recommendations are reverse-scored from 5
(strong buy) to 1 (strong sell). The consensus recommendation for a ticker equals the mean outstanding recommendation at
the end of a calendar quarter, based on a minimum of three recommendations. Firms are grouped into quintiles at the
beginning of the next quarter based on the change in consensus. Panel A reports summary statistics on the consensus change
in each quintile using the 2002 I/B/E/S tape (all “no change” observations are included in Q3). Estimates are formed once a
quarter, in cross section. We aggregate the quarterly estimates and report the time-series mean. Panel B reports the fraction of
ticker/quarter observations in each change quintile (and for the full sample) that are classified in a different quintile when
using the 2004 I/B/E/S tape. Panel C reports monthly portfolio returns for a simple trading strategy (“Spread”) that each
month buys stocks in the highest change quintile (Q5) and shorts stocks in the lowest change quintile (Q1), with a one-month
holding period. The “4-factor alpha” is the intercept from a regression of monthly portfolio spread returns on (i) the excess of
the market return over the risk-free rate, (ii) the difference between the monthly returns of a value-weighted portfolio of small
stocks and one of large stocks (SMB), (iii) the difference between the monthly returns of a value-weighted portfolio of high
book-to-market stocks and one of low book-to-market stocks (HML), and (iv) the difference between the monthly returns of
a value-weighted portfolio of high price momentum stocks and one of low price momentum stocks (UMD). The “4-factor
plus industry alpha” adds industry excess returns to the regression (i.e., value-weighted excess returns for each of ten industry
segments as defined by Kenneth French). The strategy is performed separately on the 2002 tape and the 2004 tape, and
differences between the two tapes are reported. t-statistics are in parentheses.
Panel C: Monthly Portfolio Returns (in %) from Consensus Change Quintiles (2002 tape v. 2004 tape)
Consensus
increase (Q5)
portfolio return
Consensus
decrease (Q1)
portfolio return
Spread (Q5-
Q1) in raw
returns
Spread (Q5-
Q1) in 4-
factor α
Spread (Q5-Q1)
in4-factor plus
industry α
2002 tape 1.872 0.98 0.892 0.482 0.462
(3.06) (1.53) (5.60) (3.82) (3.05)
2004 tape 1.882 0.847 1.034 0.65 0.658
(3.08) (1.32) (6.25) (4.70) (4.08)
Difference (2004 minus 2002) 0.142 0.168 0.196
(2.08) (2.14) (2.15)
% difference rel. to 2002 tape 15.9% 34.9% 42.4%
Panel A: Consensus Recommendation Change Quintiles (Change = Current – Prior) Quintile on 2002 tape Mean no. obs Mean Std dev Minimum Maximum
Best=Increase (Q5) 536.30 0.43 0.043 0.32 0.49
Q4 439.65 0.11 0.018 0.07 0.14
Q3 644.48 -0.00 0.006 -0.02 0.00
Q2 516.83 -0.14 0.032 -0.21 -0.08
Worst=Decrease (Q1) 516.04 -0.52 0.058 -0.64 -0.36
Mean of 23 quarterly samples 2653.30 -0.02 0.021 -0.09 0.02
Panel B: Migrations in Quarterly Consensus Change Quintiles (2002 tape vs. 2004 tape) 2002 tape All changes Quintile according to 2004 tape
Quintile on 2002 tape No. No. % Q5 Q4 Q3 Q2 Q1
Q5 (=increase) 12,335 1,605 13.0% 1,082 350 97 76
Q4 10,112 2,367 23.4% 1,004 689 599 75
Q3 14,823 2,436 16.4% 376 828 930 302
Q2 11,887 2,646 22.3% 103 624 878 1,041
Q1 (=decrease) 11,869 1,544 13.0% 84 58 248 1,154
All ticker-quarters 61,026 10,598 17.4% 1,567 2,592 2,165 2,780 1,494
46
Table VIII. Patterns of Changes at the Analyst-ticker Level.
The table exam
ines the patterns of the alterations, deletions, additions, and anonymizations at the analyst level. We define a “history” as a sequence of
recommendations by analyst i for ticker k, ordered in calendar tim
e. Reported percentages do not sum to 100% because histories consisting of two recommendations
are excluded in this table if only one of the tw
o records is affected by the changes.
Alterations
Deletions
Additions
Anonymizations
row
No.
%
No.
%
No.
%
No.
%
All affected records
10,698
4,923
19,204
19,904
Affected records represent:
all I/B/E/S records by the analyst
(1)
418
3.9%
382
7.8%
622
3.2%
1,210
6.1%
instances where an analyst covered a ticker only once
(2)
2,617
24.5%
543
11.0%
2,324
12.1%
2,972
14.9%
an analyst’s entire history for a ticker
(3)
1,181
11.0%
770
15.6%
2,407
12.5%
8,670
43.6%
an analyst’s history for a ticker after a certain date
(4)
975
9.1%
793
16.1%
1,584
8.2%
1,724
8.7%
an analyst’s history for a ticker before a certain date
(5)
750
7.0%
394
8.0%
2,201
11.5%
3,572
17.9%
selective records from an analyst’s history for a ticker
(6)
3,419
32.0%
1,573
32.0%
7,757
40.4%
1,017
5.1%
47
Table IX. Portfolio Performance of Anonymized and Non-Anonymized Recommendations. This table reports average daily percentage buy-and-hold abnormal returns for portfolios of upgraded buy recommendations
(which includes upgrades to buy or strong buy from a previous hold, sell, or strong sell recommendation, and initiations with
a buy or strong buy rating) and portfolios of downgraded hold/sell recommendations (which includes downgrades to hold,
sell, or strong sell from a buy or strong buy, and initiations with a hold, sell, or strong sell rating), for all anonymized
recommendations and non-anonymized recommendations by analysts who anonymized at least one of their recommendations
(but not all). Corresponding t-statistics are in parentheses. Panel A reports the results for the full sample of upgrade
recommendations, Panel B reports the results for the sub-sample of recommendations by those analysts who by 2002 had not
left the industry, and Panel C reports the results for the sub-sample of recommendations that are “bold” (i.e., they deviate in
absolute terms from the consensus by at least one notch). Panel D reports the results for the full sample of downgrade
recommendations. Columns (1) and (2) report the average daily abnormal returns for the entire sample period (October 29,
1993 to July 18, 2003); Columns (3) and (4) and columns (5) and (6) report the average daily abnormal returns for the period
through March 10, 2000 (the date of the NASDAQ market peak) and the period subsequent to March 10, 2000, respectively.
The “4-factor alpha” is the intercept from a regression of the daily portfolio excess return on (i) the excess of the market
return over the risk-free rate, (ii) the difference between the daily returns of a value-weighted portfolio of small stocks and
one of large stocks (SMB), (iii) the difference between the daily returns of a value-weighted portfolio of high book-to-market
stocks and one of low book-to-market stocks (HML), and (iv) the difference between the daily returns of a value-weighted
portfolio of high price momentum stocks and one of low price momentum stocks (UMD). The “4-factor plus industry alpha”
is the intercept from a regression of the daily portfolio excess return on (i)-(iv) plus (v) industry excess returns (value-
weighted excess returns for each of ten industry segments as defined by Kenneth French). To control for non-synchronous
trading, all regressions also include one-day lags of each of the independent variables.
10/29/1993 to 07/18/2003 10/29/1993 to 03/10/2000 03/11/2000 to 07/18/2003
4-factor 4-factor plus 4-factor 4-factor plus 4-factor 4-factor plus
α Industry α α industry α α industry α
(1) (2) (3) (4) (5) (6)
Panel A: Upgrades to buy, all analysts with at least one anonymized recommendation
Anonymized (a) -0.0107 -0.0100 0.0001 -0.0006 -0.0269 -0.0243
(-1.78) (-1.71) (0.02) (-0.09) (-2.09) (-1.90)
Non-Anonymized (b) 0.0043 0.0042 0.0100 0.0084 -0.0033 0.0017
(0.98) (1.00) (2.29) (2.12) (-0.34) (0.18)
(b) – (a) 0.0149 0.0142 0.0099 0.0089 0.0236 0.0260
(2.97) (2.87) (1.96) (1.78) (2.12) (2.38)
Panel B: Upgrades to buy, all analysts with at least one anonymized recommendation (excluding possible retirees)
Anonymized (a) -0.0083 -0.0099 0.0006 -0.0020 -0.0214 -0.0207
(-1.40) (-1.72) (0.10) (-0.36) (-1.69) (-1.65)
Non-Anonymized (b) 0.0074 0.0053 0.0104 0.0076 0.0053 0.0085
(1.60) (1.21) (2.18) (1.83) (0.54) (0.90)
(b) – (a) 0.0157 0.0152 0.0098 0.0097 0.0267 0.0292
(3.10) (3.06) (1.95) (1.94) (2.36) (2.65)
Panel C: “Bold” upgrades to buy, all analysts who anonymized at least one recommendation
Anonymized (a) -0.0257 -0.0264 -0.0191 -0.0191 -0.0419 -0.0474
(-2.89) (-3.01) (-2.12) (-2.15) (-2.16) (-2.45)
Non-Anonymized (b) -0.0052 -0.0089 -0.0059 -0.0066 -0.0059 -0.0121
(-1.01) (-1.86) (-1.13) (-1.40) (-0.55) (-1.19)
(b) – (a) 0.0206 0.0176 0.0132 0.0125 0.0360 0.0354
(2.35) (2.01) (1.49) (1.41) (1.90) (1.84)
Panel D: Downgrades to hold/sell, all analysts who anonymized at least one recommendation
Anonymized (a) -0.0067 -0.0132 -0.0079 -0.0121 -0.0160 -0.0236
(-1.04) (-2.18) (-1.26) (-2.06) (-1.18) (-1.78)
Non-Anonymized (b) -0.0029 -0.0061 -0.0071 -0.0079 -0.0041 -0.0092
(-0.68) (-1.50) (-1.88) (-2.29) (-0.42) (-0.98)
(b) – (a) 0.0038 0.0072 0.0008 0.0042 0.0119 0.0144
(0.72) (1.39) (0.14) (0.80) (1.08) (1.30)
48
Table X. The Effect of Alterations, Deletions, Additions, and Anonymizations on Subsequent Career Outcomes, 2003-2005.
This table estim
ates the effect of alterations, deletions, additions, and anonymizations on analysts’ subsequent career outcomes. The unit of analysis is an analyst-year. W
e
focus on analysts on the 2002 I/B/E/S tape; analysts who entered the profession after 2002 are ignored. The dependent variable equals one if the analyst that year changed
jobs (col. (1)), moved to a ‘large’ brokerage firm
employing 25 or more analysts (cols. (2) and (3)), moved to a ‘sm
all’ brokerage firm
with fewer than 25 analysts (cols.
(4) and (5)), exited the industry (i.e., ceased to contribute research to I/B/E/S the following year; col. (6)), or was rated the top stock picker in her sector in the following
year’s Wall Street Journal “B
est on the Street” survey (col. (7)). The sample in col. (7) is restricted to analysts who are eligible for the WSJ survey, which we approxim
ate
as analysts who cover a m
inim
um of five stocks in the relevant calendar year. Relative forecast accuracy and boldness are defined as the analyst’s scaled rank (relative to
other analysts covering the same stocks) of deviations between forecast and subsequent earnings realization and of the absolute deviation from the earnings forecast
consensus, respectively, while relative forecast optimism is defined using a dummy variable = 1 if the analyst’s forecast for a stock exceeds consensus. Each m
easure is
averaged across stocks she covers in years t-2 through year t; see Hong, Kubik, and Solomon (2000) and Hong and Kubik (2003). Other explanatory variables are defined
in T
ables II-IV. All m
odels are estimated using probit. To conserve space, intercepts and year fixed effects are not shown. Results are robust to including random
brokerage firm
effects to control for otherwise omitted heterogeneity arising from differences across brokerage firm
s. Standard errors, shown in italics, are clustered by
analyst (i.e., observations are assumed to be independent across analysts but not necessarily w
ithin). W
e use ***, **, and * to denote significance at the 0.1%, 1%, and 5%
level (tw
o-sided), respectively.
49
Table X. Continued.
Changes
Promotion: Moves to large
brokerage firm
from …
Demotion: Moves to small
brokerage firm
from …
Exits
Becomes
WSJ top
jobs
…
any firm
… small firm
… any firm
…
small firm
industry
stock picker
(1)
(2)
(3)
(4)
(5)
(6)
(7)
=1 if analyst’s recommendations were …
… altered
0.058
-0.035
0.062
0.128*
0.188*
0.009
0.084
0.054
0.070
0.119
0.065
0.083
0.057
0.106
… dropped
-0.014
-0.089
-0.224
0.056
0.104
0.032
0.192*
0.056
0.075
0.133
0.065
0.089
0.056
0.095
… added
0.104*
0.181***
0.243*
0.009
0.035
0.076
0.172*
0.046
0.057
0.101
0.058
0.079
0.048
0.087
… anonymized
0.159***
0.254***
0.305**
0.033
0.036
-0.008
0.170
0.050
0.062
0.112
0.061
0.079
0.052
0.093
=1 if analyst is ranked II all-star
-0.740***
-0.387***
-1.400***
-1.385***
-0.658***
-0.091
0.098
0.104
0.308
0.313
0.093
0.115
=1 if analyst is top stock picker
-0.271
-0.289
-0.178
-0.158
-0.088
-1.045**
0.187
0.251
0.451
0.231
0.310
0.382
seniority (log years in I/B/E/S)
-0.083*
-0.214***
-0.230**
0.032
-0.007
0.135***
-0.040
0.034
0.045
0.080
0.040
0.054
0.033
0.069
relative forecast accuracy
-0.026
0.278
0.473
-0.198
-0.140
-0.359*
0.556
0.170
0.227
0.400
0.196
0.237
0.156
0.371
relative forecast optimism
0.165
-0.019
0.053
0.227
0.282
-0.041
-0.216
0.107
0.141
0.256
0.126
0.159
0.099
0.255
relative boldness
-0.596***
-0.570*
-0.295
-0.451*
-0.362
-0.182
0.558
0.178
0.226
0.400
0.211
0.259
0.171
0.366
ln(analyst’s no. of stocks covered)
-0.157***
0.030
0.067
-0.262***
-0.359***
-0.914***
0.325***
0.033
0.044
0.083
0.037
0.049
0.034
0.100
ln(no. of analysts covering same stocks)
0.214***
0.274***
0.427***
0.138**
0.244***
0.266***
-0.294***
0.039
0.051
0.084
0.046
0.067
0.040
0.075
= if analyst worked at sm
all broker in t-1
-0.104
0.277***
0.055
0.047
Diagnostics
Pseudo R
2
3.8 %
3.8 %
6.3 %
6.2 %
8.5 %
17.0 %
3.7 %
Wald test: all coefficients=0 (χ2)
166.2
***
109.4
***
60.6
***
159.1
***
117.7
***
876.5
***
44.2
***
Number of observations
7,696
7,696
2,468
7,696
5,228
7,696
6,804
Recommended